GEO vs SEO: What Changes and How to Win AI Traffic (2026)

GEO vs SEO: what GEO is and what changes

With SEO, a page that ranks well in Google or Bing gets visits without paying for ads, as long as it keeps its place: these unpaid results and visits are called organic. SEO work comes in four kinds, as the U.S. Chamber of Commerce listed them in August 2026, on its site for small businesses:

• On-page SEO: the words, titles and structure of each page, chosen for what people search.
• Technical SEO: a site that loads fast, has no broken links, and that search engines can read and add to their index, the database of pages they show results from.
• Off-page SEO: what happens on other sites, above all the links that point to your pages, mentions of your brand and reviews.
• Local SEO: showing up in searches for a place, such as "dentist near me", with the same name, address and phone number everywhere and a Google Business Profile, the free listing that shows a business, its hours and its reviews in Google Search and Maps.

GEO aims somewhere else: the answer an AI writes, rather than a place in the results. When a buyer asks ChatGPT "which payroll software should a 20-person company use?", the AI writes a few paragraphs, names a handful of brands and sometimes adds links to the pages it drew on.

The main engines in question are ChatGPT, Perplexity (an answer engine that shows its sources), Google's Gemini, Microsoft's Copilot, Anthropic's Claude, and two AI features inside Google Search. AI Overviews is the summary above the results, which Google rolled out to everyone in the US in May 2024. AI Mode is a tab where you chat with the search engine.

The same work goes by several names: GEO, AEO (answer engine optimization), LLMO (large language model optimization, a large language model being the program that writes the answers), "AI SEO" or "AI search optimization". Wikipedia notes there is no agreed definition. GEO in this sense has nothing to do with geography: it isn't geo-targeting, which shows content by location, or local SEO.

An AI answer can show your brand in three ways, and the difference matters for everything that follows:

• A citation: one of your pages appears among the sources of the answer, usually with a link.
• A mention: the answer names your brand in its text, with or without a link.
• A recommendation: the answer names your brand as the one to choose.

A citation doesn't bring the other two with it. Devesh Khanal and Benji Hyam run Grow and Convert, an SEO and GEO agency. The agency also sells its own AI-visibility tracker, a tool that counts how often AI answers name or cite a brand. In September 2026, on their YouTube show, they showed what their tracker recorded for a former client, Level AI, which sells AI software for call centers.

On a newer topic for the company, AI call monitoring, Level AI's own site was the source the AI engines cited most, level with one other site. Yet the brand came about tenth among the brands the engines named. In Grow and Convert's reading, the engines used Level AI's pages without yet linking the brand to that topic.

Your own pages on the topic are the start, and Level AI already had them. What was missing, in Grow and Convert's view, was other sites: reviews and the "best tools" lists the engines already cite, describing the brand for that topic, the same way everywhere.

Their example: if 20 different sites name Level AI as a company that does AI call monitoring, and describe it the same way, the AIs get one clear message. "The more consensus there is", the better.

Graphite, a growth agency, found the same gap in nearly half a million AI shopping answers. Cassie Clark, an AI-visibility consultant, reported its study on her podcast Found in AI in October 2026. Large and small retailers each made up about 38% of the stores cited. Yet large retailers were the top recommendation about two and a half times as often as small ones.

SEO vs GEO, point by point

Set next to each other, the two differ from what you aim for to what decides who shows up:

What works for both SEO and GEO

Much of the work counts twice: Google's AI features rest on the same ranking and quality systems as its ordinary results, its July 2026 guide for site owners says. Five practices serve your rankings and your citations at once:

1. Pages search engines can read: indexed, with their text in the code the site sends rather than added later by JavaScript, code the browser runs after the page arrives, so that Google, Bing and the AIs can all read them.
2. What only you know: your own data, tests, prices and customer case studies, rather than what every other page already says.
3. Passages that stand on their own: each passage names its subject in its first sentence and makes sense when an AI quotes it alone.
4. A brand other sites talk about: links and mentions on other sites help your rankings, and they help an AI name you.
5. Pages kept current: Microsoft recommends regular updates, so that its AI answers cite the latest version of a page (February 2026). Newer isn't automatically better, though. In an August 2026 study by Rankability, a company that sells SEO software, half the pages AIs cited were under 61 days old, going by their dates, and so were half the pages in search results. So update a page when its facts change, not to make it look new.

To see why ranking still comes first, and what it doesn't settle, you need to know how an AI picks the pages it reads.

How an AI chooses the pages it cites

An AI answer comes from two places. The first is what the AI's model, the program that writes its answers, learned in training, from a large share of the web, up to a cutoff date. The second is a web search run at the moment of the question, whose results the model reads before writing (people call this grounding).

Anything newer than the training cutoff can only come from that live search. Answers with links usually come from that search, and answers with no link at all mostly come from what the model learned in training.

The question becomes a few short searches

People don't type the same thing into ChatGPT as into Google. They describe their situation in a few sentences. The AI then turns that message into a few short searches that it sends to a search engine. This is called query fan-out, and the short searches are often called sub-queries.

Google describes it for its own AI features in its guide for site owners, "Optimizing your website for generative AI features on Google Search", updated in July 2026. One question becomes several searches: a question like "how to fix a lawn that's full of weeds" fans out into "best herbicides for lawns", "remove weeds without chemicals" and "how to prevent weeds in lawn".

OpenAI's help center says the same of ChatGPT: it "typically rewrites your query into one or more targeted queries" sent to its search partners.

Benjamin Tannenbaum, who founded AISO, a tool that collects these sub-queries, described the pattern in June 2026 on the James Dooley Podcast, the show of James Dooley, a UK SEO consultant. Someone tells ChatGPT they're visiting New York, want a pepperoni pizza and can't stand spicy food. The AI searches something like "pepperoni not spicy New York 2026".

When the person doesn't name a place, the AI adds the one it detects from their connection, and it often adds the year. For a local business, that means the short searches usually contain the city. In Benjamin Tannenbaum's view, a page whose key content is brought up to date for the current year is more likely to match those dated searches, as long as it ranks for them.

In his data, ChatGPT runs about two searches per question, Gemini about ten. His explanation is cost: OpenAI pays a search provider for each search, while Google's own searches cost it far less.

Two things follow for you. Many long, personal questions end up as the same few short searches. And your page has to rank on those short searches, not on the question as the person typed it.

Which search engine the AI asks, then how it chooses

Google's AI Overviews and AI Mode search Google's own index, the database of pages Google has read. Google's guide says these features "are rooted in our core Search ranking and quality systems". OpenAI names no search partner in its help pages.

Practitioners disagree on what ChatGPT uses. Dan Petrovic, who runs DEJAN, an Australian SEO agency, said on the podcast of Ahrefs, a company that sells SEO software, in August 2026 that his tests point mostly to Bing for ChatGPT, and to Brave, an independent search engine, for Claude. He admits these are "black box systems", which nobody outside can see into.

Nathan Gotch, cofounder of Rankability, said on the podcast Search with Candour in October 2026 that his company's data found no close match: ChatGPT's sources didn't follow Bing's, Google's or Brave's results. He thinks OpenAI is building its own index.

Rankability's August 2026 report also cites an earlier study by the company, of 2,824 AI citations. It put a number on that mismatch: 77% of the pages ChatGPT cited didn't appear at all in the search results Rankability measured. For Perplexity, the share was 12.2%.

Studies like these compare an AI's sources with the search results for the question as it was asked. Rankability's August 2026 study typed the same 31 questions, one per topic, into the AI engines and into search engines.

Ahrefs did the same in July 2025 with 15,000 prompts, the requests people type to an AI. It found the same gap: only 12% of the pages the AIs cited were in Google's top ten for the prompt, about 8% for ChatGPT and 28.6% for Perplexity.

Ahrefs' explanation is query fan-out: the AI may never search the question itself, only shorter variants of it, and it cites pages that rank for those. Part of the gap, then, comes from measuring the wrong searches.

Once it has its search results, the model sorts them its own way. Dan Petrovic asked Gemini a million questions to map its preferences. Gemini, he found, comes with favorite brands from its training, and others it almost never puts forward. BYD, the Chinese carmaker and one of the world's largest, wasn't among the hundred brands Gemini favored most across those million questions. Chinese brands came out poorly in general, in his reading.

Those favorite brands can make the AI put the ninth search result first, in his view.

Graphite saw a related effect in ChatGPT in October 2026: ChatGPT adds brand names to its own searches. Across about 957,000 sub-queries for questions that named no brand, 59.6% contained one. A test published on this site, in its comparison of AI visibility tools, found the same on 26 October 2026. Asked 40 times which GEO software to use, ChatGPT used without an account named a tool in 60 of its 80 sub-queries, although the question named none.

In other words, the AI searches for the brands it already knows, and a brand it has never heard of may not even be searched for.

Benjamin Tannenbaum adds a detail you can't check yourself, from studies he says he runs for clients: ChatGPT often answers from the search results' titles and short descriptions without opening the pages. Treat it as one practitioner's claim. If he's right, your page's title and description, the text shown for it in search results, have to answer the question on their own.

So two things decide whether you show up. First, for the AI to cite your pages, ranking on the short searches it runs is the surest way in, above all for Google's AI features, which search Google's own index. Second, for the AI to name your brand when the question names none, the brand has to be known already: named around the web before the question is ever asked.

How much ranking helps with ChatGPT is less clear. ChatGPT's sources match the rankings for the question as typed much less closely. Part of the reason is that the question as typed is the wrong search to compare, and ChatGPT doesn't show its short searches on screen.

GEO and SEO: why SEO comes first, and where it stops

Since the AI reads pages from the results of the searches it runs, your place in those results decides a large part of your chances. Google's July 2026 guide asks the question itself: "Is SEO still relevant for generative AI search? In short, yes!"

The Rankability study measured how far that goes. In August 2026, the company compared 1,645 AI citations with 916 ordinary search results on 31 topics, across 11 AI engines, from ChatGPT and Claude to Google's AI Overviews and AI Mode:

• Ranking gets you considered. Pages in the top three of ordinary search results (Google, Bing and other engines) were cited by at least one of the 11 AI engines 98.9% of the time. For pages in the top ten, the figure was 90%, and for pages beyond the top ten, 49.3%. With 11 engines, "at least one" is a low bar: the study doesn't say how often a single engine, such as ChatGPT, cited them.
• Ranking doesn't decide which pages the AI cites first. On each topic, Rankability took the top ten search results and the first ten pages the AI engines cited, all engines together: 310 pages on each side over the 31 topics. About 139 pages were on both lists, 44.8% of each. The other 171 citations, 55.2%, came from elsewhere: 63 ranked beyond the top ten, and 108 weren't in the measured results at all.

The two figures count different things. Almost every page that ranks near the top gets cited somewhere, yet more than half of the pages cited first don't rank in the top ten for the question as typed. Rankability notes that these are correlations, not causes.

The pages the AI cites may also rank for its own short searches, which these comparisons didn't check, so the real overlap may be larger. Read together, the two findings say that ranking is the main way into the pool the AI picks from, not the only one, and that something else decides who gets picked.

Practitioners who have done SEO for years read it the same way. Lily Ray, vice president of SEO and AI search at the US agency Amsive, asked on the podcast Muscles of Search in April 2026 that people check who is talking first: "Are they selling a product that benefits them if GEO is totally different than SEO? The answer is almost always yes."

Charlie Whitworth, a former journalist with fifteen years in SEO who founded the UK agency Whitworth, told Search with Candour, an SEO podcast, in October 2026: "GEO is not a separate channel. It's a layer of organic search."

His agency's main case study of 2026 involved cleaning up a client's site. The agency removed the old pages nobody needed, merged pages that competed for the same subject, and stopped search engines' crawlers, the programs that read pages, from reading near-duplicates, pages that say almost the same thing. Visits sent by AI assistants rose with Google visits "almost immediately… the trend lines are almost identical", he said, without giving the figures.

I run this site and a few others, and I don't treat GEO as a separate channel either. I don't split my time between SEO and GEO: working on their SEO is working on their GEO.

Google itself takes the same line. In June 2026, Brendon Kraham, a vice president at Google, summed up its position in an article titled "Good SEO is good GEO". The same month, Cassie Clark, who sells AI-visibility consulting herself, answered that article on her podcast, Found in AI. For her, the formula holds for Google's AI features, which are built on Google's own ranking.

It holds less, in her view, for ChatGPT, Perplexity or Copilot, which don't rely on Google's index. Rankability's figures back Cassie Clark for ChatGPT, whose sources matched search rankings far less than the other engines'. They fit Perplexity less well: Perplexity doesn't use Google's index either, yet its top sources were closer to Google's top results than any other engine's.

Cassie Clark sums up the difference in her own words: SEO is about getting indexed, GEO is about getting chosen. In other words, SEO gets your pages into the pool the AI reads, and GEO gets your brand picked from it.

Nathan Gotch, the Rankability cofounder, sees the difference in one of his own sites. He sells an SEO course. To show his students how a site climbs on Google, he built a site of a single page. Its address is the search it aimed at, word for word: "St. Louis SEO consultant", what SEOs call an exact-match domain.

The site got good links from other sites and nothing else: no Google Business Profile and no reviews. It held first place on Google for three to five years.

"If you just do that, you will not win in AI", he said, except in a market with almost no competition. In a market as ordinary as local chiropractors, where competitors have some 50 reviews each, those reviews already beat a page that only has links in AI answers, in his view. "With SEO, you didn't need a brand at all." With GEO, you do.

Google, Charlie Whitworth, Cassie Clark and Nathan Gotch all agree on the order of the work. SEO gets your pages into the results the AI reads. Getting your brand named is extra work on top, largely on other sites. Which of the two weighs more, your place in search results or what others say about you? My answer: nobody can say today. There is a great deal of research on it and no certainty yet.

How to win AI traffic: the method, step by step

AI traffic comes from being cited: someone clicks your page in an answer. Being named brings visits too, when a buyer looks you up after reading an answer, though those visits arrive through a Google search or with your address typed in, so your analytics won't show them as AI traffic.

The method below aims at both, in an order drawn from what the practitioners quoted here do. Each step has its proof, the sign that it's done, and its usual trap.

Let the AI crawlers in

Before anything else, the AI's crawlers have to reach your pages. A crawler, or bot, is the program a search engine or an AI sends to read web pages. Each site tells crawlers what they may read in a small file called robots.txt. Most checks below are settings: ask your AI agent, if you build with one, or the service that runs your site to check each of them.

OpenAI lists four crawlers (as of 29 October 2026), Anthropic three and Perplexity two. OpenAI and Anthropic each keep the crawler that feeds their search apart from the one that collects pages to train their models, and blocking one doesn't block the other.

On the sites I run, I let every AI crawler in, training crawlers included. The search crawlers help AIs find my pages when someone asks a question. With some luck, the training crawlers get those pages into the data a future model learns from. That model can then answer with what it learned from my pages, without running a search. To me, that is worth a great deal.

All of these get in unless something stops them: your robots.txt for most of them, your firewall for all of them. A firewall is the filter in front of your site that turns away unwanted traffic, and it can block AI crawlers without anyone choosing to.

Cloudflare, a service that sits in front of websites to serve them faster and filter their traffic, shows how that can happen. Since 15 September 2026, when you add a site to Cloudflare, it offers you ready-made settings for AI crawlers. These settings differ depending on whether your site earns money from ads.

Cloudflare's reason is that an ad pays only when a person sees the page. An AI trained on your pages gives its answer without anyone coming to your site. When an AI reads a page on someone's behalf, that person doesn't see the ads either. For a site with ads, the settings offered are therefore stricter.

Without ads, the settings offered let every crawler in. With ads, they change with what each crawler does:

• Crawlers that only train models, such as GPTBot or ClaudeBot, are blocked on your whole site.
• Crawlers that read a page at a person's request are blocked on the pages that show an ad. ChatGPT-User, for example, reads a page when someone asks ChatGPT a question.
• Search crawlers can still read your pages. They make your site appear in AI search, as OAI-SearchBot does in ChatGPT's search.

Googlebot and Bingbot, the crawlers of Google and Bing, are a special case: what each one reads serves both search and the training of AI models. Blocking them would risk removing your site from Google and Bing. For these two, the settings offered to sites with ads block nothing.

Instead, these settings add an instruction to your robots.txt, asking that no AI model be trained on your pages: Cloudflare calls this choice "Disallow AI Training". Googlebot and Bingbot keep reading your pages for search.

The trap is in the same Cloudflare settings, which you can change when you add your site or later. For training, they offer two other choices: "Block" (on your whole site) and "Block on pages with ads". Since 15 September 2026, both also stop Googlebot and Bingbot from reading your pages for search. If you pick one to refuse training, Google and Bing stop reading the pages concerned.

Charlie Whitworth, who founded the UK SEO agency Whitworth, saw this trap coming. Many sites, he said, may have switched on a block of AI training bots long ago. After the change, such a site "could actually block Googlebot inadvertently". He said it on Search with Candour, an SEO podcast, in an episode recorded before 15 September and published in October 2026.

Your host, the service that puts your site online, can tell you whether your site runs through Cloudflare. If it does, have your host or your AI agent check these settings. They are in Cloudflare's dashboard, under Security Settings, "Configure AI bot policies" (Cloudflare's documentation, as read on 1 November 2026).

Charlie Whitworth also advises checking which crawlers actually reach which pages. Your server, the computer your site runs on, keeps a record of every visit in its logs. Ask your host or your AI agent to read them for you.

On Google's side, a page itself can appear in AI Overviews and AI Mode on two conditions. It has to be in Google's index. Google also has to be able to show a snippet of it, the short text under the title in the results.

By default, Google can do both: you are the one who can forbid it, with an instruction written in the page's code for search engines, such as noindex (keep this page out of the index) or nosnippet (show no snippet of it).

Even without such an instruction, Google doesn't index every page: it decides for each one whether to keep it in its index. To find out whether one of your pages is there, use Search Console, Google's tool for site owners (free, as checked on 29 October 2026). Add your site to Search Console first. Then paste the page's address into its URL Inspection tool: it tells you whether Google has kept the page in its index.

For the whole site, there is also a setting. Since 31 August 2026, Search Console has let you include or exclude your site from AI Overviews and AI Mode. Sites are included by default. Excluding takes effect in a day or two and doesn't change your place in ordinary results: it's meant for publishers who don't want their text in those answers. If you want to be cited, leave it on "include".

On Bing's side, IndexNow is a way to tell the search engine the moment a page changes, instead of waiting for its next visit. Bing recommends it so that its AI answers cite the current version of your pages. Bing and a few other search engines accept these notices. Google doesn't.

If ChatGPT does lean on Bing, IndexNow helps there too. A host or an AI agent can switch it on for you.

The text of each page should also be in the code the site sends, without waiting for JavaScript, the code that runs in the browser after the page loads.

Will Critchlow, who heads SearchPilot, an SEO testing platform, said on the podcast Outperform in March 2026 that "a lot of the new LLM crawlers don't execute JavaScript at all". In Rankability's August 2026 study, 92.5% of the pages AI engines cited showed their main text without JavaScript, a correlation that doesn't prove the cause.

On finngarden, the site you're reading, the robots.txt lets in the crawlers of OpenAI, Anthropic and Perplexity by name, one by one. It also lets Google use its pages for Gemini, through Google-Extended. Since 29 October 2026, each time a new version of the site goes live, IndexNow is told which pages changed.

Each article also carries structured data, labels invisible to readers that tell search engines who wrote the page and when. I always add it, on every site I run. Its job is to help search engines read each page and qualify it for rich results in Google, the listings with extra details, not to win a place in AI answers.

Since August 2026, the site has published an llms.txt file, a plain-text summary meant for AI tools, and an llms-full.txt file with the text of every page.

It also offers an access point built on the Model Context Protocol (MCP), a standard way for AI agents, AIs that act on someone's behalf, to connect to a service. Through it, an agent can search and read the site's pages directly. This access point is an experiment: I added it to see whether agents would use it.

Since 10 September 2026, the site's pages have also offered WebMCP tools. WebMCP is a proposed browser feature, drafted by engineers at Google and Microsoft and not yet a web standard. A page uses it to tell an AI agent working in the browser which actions the page offers, each described in plain words.

On finngarden, an agent can use these tools to search and read the site. It can also fill in the newsletter sign-up form for the person it works for, who then confirms with a click.

Of the MCP access point and the WebMCP tools, WebMCP is the one I believe in, and I have put it on almost all my sites. It lets agents act on a site, not only read it, and I expect that use to grow.

OpenAI has started to use it, and Google checks for it. Since 31 August 2026, ChatGPT's desktop app can use the WebMCP tools of the sites that offer them, in its built-in browser, with a supported account and model, OpenAI's release notes say.

Lighthouse, Google's free audit tool built into Chrome, has an Agentic Browsing category, a set of checks for sites that AI agents use. Among them, it looks for a site's WebMCP tools and its llms.txt file (as of 3 November 2026).

These files and tools serve the AI agents and developers who come to read or use the site. No evidence shows they get a site cited in AI search, and Google says it doesn't use llms.txt files.

Choose your questions and read the searches behind them

Once the crawlers can reach your pages, start from what your buyers really ask. Kevin Indig, a growth advisor who has worked on search with Airbnb, Shopify, Ramp and G2, recommended in a July 2026 webinar picking 10 to 20 questions about your category and your product, found in your sales calls, to ask the AIs and track.

His method: take 30 to 50 recorded calls, with the callers' consent, as the law and your contracts require, and ask an AI what problems the prospects describe, what they use today and which alternatives they compare. If you don't record calls, your support emails, chats and reviews hold the same questions.

He followed the same logic in 2024 for Ramp, a US company that sells corporate cards and expense software, asking the people directly: "We asked prospects, accountants, bookkeepers… what do you use right now? What alternatives do you prefer?" The articles written from those answers "performed incredibly well", he said.

Don't look for the search volume of a question people ask an AI, the number of times it's typed each month. Devesh Khanal of Grow and Convert explains why the tools that sell these counts, called "prompt volumes", can't be trusted. Their data comes from people paid or rewarded to have their browsing tracked, and the buyers you want, such as the technology heads of large companies, are rarely among them.

And as Dan Petrovic points out, two people seldom type the same question. He estimates demand another way: he adds up the search volumes of the short searches the AI actually runs, which any SEO tool gives, as does Google's Keyword Planner, the free search-volume tool in Google Ads.

Then work out those short searches. In ChatGPT, run each question with web search on. If ChatGPT answers without searching, OpenAI's help explains how to force it: in the message box, select "View all tools", then "Search", or type "/" and choose "Search".

On 29 October 2026, OpenAI's help said web search worked on the free plan and even without an account. The answer then lists the sites it drew on. OpenAI's help doesn't show where to read the short searches themselves. Benjamin Tannenbaum describes two ways: in the data your own browser receives while ChatGPT answers, which your AI agent can read for you, or with a free browser extension that reads them for you.

Four things work with no technical help. Note, every week, the sources ChatGPT lists for each question. Write down the short searches each question most likely becomes, the way Google's example turns the lawn question into "best herbicides for lawns". Most are the product or the need, with a city or the year added when they matter.

And open the AI Performance report of Bing Webmaster Tools, Microsoft's tool for site owners (free, as checked on 29 October 2026). It lists a sample of the real searches Microsoft's AI ran before citing your pages. That report has two limits: it covers only Microsoft's AI, such as Copilot, and it shows nothing until your pages are being cited.

Last, type each short search into Google and Bing and look at the first results. If they are pages like the one you could publish (a case study, a guide, a comparison, a page about your offer), your site can aim for the top.

If they are "best of" lists, review pages, media articles or forum threads published by others, your page will struggle to pass them. The work there is to be named in those pages.

Benjamin Tannenbaum ranks these results by effort. Your own page is the cheapest to get to the top, sometimes with a single edit, in his experience. A media list costs more to get into, but one media site often ranks two or three times for the AI's searches, so one mention there counts several times. A thread on Reddit, the large forum site, is the riskiest, because its members spot and punish a seller posing as a user.

On your site: pages an AI can use

For your own site, Google gives the rule in its July 2026 guide: publish "non-commodity" content, content that isn't the same as everyone else's. Its example: an article called "7 Tips for First-Time Homebuyers" repeats what everyone knows. A buyer's account of why they skipped the home inspection, what they saved and what they then found in the pipes brings an experience nobody else has.

Martin Splitt, of Google's team in charge of relations with site owners, said it his way on Google's podcast Search Off the Record in May 2026. Rewording a product's spec sheet adds nothing, because a chatbot can do it too. What counts is what you have tested yourself, because the AI hasn't.

For my own pages, one rule comes before all the others: answer what the person came for. A page has to give them what they expect or what they need. It is the rule a product people buy has to meet too, and it holds on Google as in AI answers.

In its July 2026 guide, Google also writes that you don't need to rewrite your pages "for AI" or cut them into small chunks: its systems understand synonyms and can pick out the right part of a page, so you don't need to repeat the exact words people search. Readers and AIs both often take a single passage, though, so the rules below are writing rules for your readers that also make your pages easy to cite:

• Say what the page is about at once. In Rankability's study, pages that introduced their subject within the first 100 words had a 22.1% chance of being among the AI's first ten citations, against 8.7% for the others.
• Make each sentence stand on its own. Repeat the subject rather than "it" or "this", give a figure rather than a vague word, and cut the filler: a passage the AI lifts alone has to make sense alone.
• Publish your own data. First-hand data is new information the AI can't find anywhere else. Kevin Indig's example: a company selling CRM software, the software businesses use to manage their customer relationships, may get press with a study on employee retention, "but that does not help LLMs recommend you when it comes to CRM". A study of how companies use CRM software does, when it answers the exact questions buyers ask.
• Keep the data dry and the story lively. In Kevin Indig's words: "When it comes to the data, you need to be dry like a scientist. When it comes to the storytelling, you can be a full-blown marketer."
• Show your prices. In a study of sites that sell to other businesses, which Kevin Indig was about to publish in July 2026, pricing was the information AI assistants found hardest when browsing for a buyer, because it was hidden behind a form or missing. A young company that wants to beat an established one can simply show its prices.
• Date your key pages and keep them current. Cassie Clark, on her podcast in June 2026, advised revisiting your most important pages to refresh their data, and saying when each figure was checked with an "as of" date. In her view, that date tells the AI the page still reflects the current situation.

Some page types do better than others. Ross Hudgens founded Siege Media, a content marketing agency, and has worked in search for about 14 years. On The Edward Show, a podcast, in September 2026, he said that his clients with the most AI traffic were those with the most "versus" pages, your brand against one competitor.

He found it by comparing his clients. His explanation: near the end of their choice, buyers ask the AI to compare two products. He builds these versus pages as pages meant to turn visitors into customers, apart from the blog, with the comparison table near the top, written by hand rather than from a template.

The competitor's prices have to be exactly right. On the show, the risk was put plainly: get them wrong and the competitor may demand that you remove the page. And "LLMs will use that as context too": the AIs may repeat your wrong prices. His advice on tone: admit the competitor is the better fit for some buyers.

Pages for people choosing between products, rather than learning about a subject, also hold up better against AI Overviews. SignLab makes ASL Bloom, an app to learn American Sign Language (ASL) and other sign languages, "Duolingo, but for sign languages". Andre, who runs SignLab, told the story on The Edward Show in July 2026.

SignLab had built a free online dictionary: one page per sign, with a video of the sign, example sentences and an invitation to download the app. A typical visitor: a parent of a deaf toddler who searched how to sign "stop". In 2025, Google's AI Overviews began showing those videos right in the results.

SignLab's rankings held, but "we lost 50% of our traffic", and app downloads from Google fell further still. A parent who sees the sign in Google has their answer, while a parent on the page sees how much more there is to learn. Moving thousands of clips to YouTube, whose videos Google shows in its results, brought back some visibility, "but I won't oversell like how powerful it was".

What brought the downloads back was pages for people choosing an app. Before, SignLab targeted searches like "how to sign hello in ASL". After, it targeted searches like "does Duolingo have ASL", and its article answering that one became its best page in the US for downloads.

"Top of funnel content is the easiest to just regurgitate for AI", Andre said, meaning the pages for people who are only starting to look into a subject. SignLab never won back the old dictionary traffic.

AIs are also used to decide between offers, and then they read your proof. James Dooley, a UK SEO consultant, told this story on his podcast Fatrank in September 2026, as a case study of his own work for Soft Surfaces, a British company that builds sports fields.

A school in Manchester wanted an artificial soccer field (a football pitch, in British English) and had four quotes. Soft Surfaces was the second most expensive. The head teacher, the school's principal, gave the four quotes to ChatGPT and asked which offered the best value.

ChatGPT noticed what the school's business manager had missed: a cheaper quote left out the asphalt base the specification required. It questioned a competitor offering a ten-year guarantee after four years in business. Then it read the proof on Soft Surfaces' site: the company's accreditations, its insurance certificate, its twenty years in business and its similar fields nearby. ChatGPT named it the best choice.

The contract was worth £572,000. The head teacher later told James Dooley the school would not have chosen Soft Surfaces if ChatGPT hadn't named it the best value.

James Dooley's lesson: when an AI is asked to choose between offers it already has, it reads the proof on your own site, from guarantees and accreditations to years in business and case studies. Which brands an AI names when nobody hands it quotes depends more, in his view, on what other sites say about them.

Off your site: get named by others

For a known brand, AIs cite what others say about it more than what it says about itself. The clearest numbers date from 2025, more than six months ago, and the 2026 accounts in this guide, from Grow and Convert to Nathan Gotch, describe the same pattern. A smaller brand can see its own pages cited, as Level AI did, yet be named only once other sites name it.

University of Toronto researchers measured it then: for the well-known brands in their study, 93.5% of the sources ChatGPT cited were media and independent sites, and 6.5% the brands' own sites.

Ahrefs compared 75,000 brands in 2025: those that appear most in AI answers are those mentioned most on the web, especially in YouTube videos, while the number of links to their site counted much less. Ahrefs notes these are correlations, not proof of cause.

The two ways in are reviews and other sites' lists.

Nathan Gotch puts it as "reviews, reviews, reviews", on the platforms the AI actually reads. For a local business, the Google Business Profile feeds Google's AI. ChatGPT, in his experience, doesn't use it and draws on Yelp, so Yelp comes second.

In the US, each trade has its own review sites: Avvo for lawyers, Angi and Thumbtack for plumbers, G2, Capterra and Trustpilot for software. He picks the three to five platforms that keep appearing among the sources the AI cites, and treats each profile "like it's your website".

For the "best tools" lists the engines already cite, he works at scale. For a category he wants to win, he runs many questions on every AI engine, collects every cited page plus the top 30 search results for each question, and removes duplicates. He ends up with "500 to 1,000 opportunities" he contacts one by one, asking each author to consider the brand.

Google sets a limit in its July 2026 guide: seeking "inauthentic mentions", bought in bulk or written by fake customers, doesn't help its AI answers, because such mentions go through the same spam filters as its results. So earn the place with facts the author can check.

Social platforms count differently from one AI to the next. In the Toronto study, ChatGPT cited no social platform for well-known brands, while Perplexity took almost a quarter of its sources from them. So whether posting on social platforms helps depends on the AI your buyers use.

Measure your visibility in AI answers

Before you measure, know that an AI almost never gives the same list of brands twice. Rand Fishkin, founder of SparkToro, an audience-research company, had 600 volunteers ask ChatGPT, Claude and Google's AI (AI Overviews, or AI Mode when no Overview showed) the same twelve questions, 2,961 times in all, in January 2026.

For ChatGPT and Google, two answers to the same question had less than one chance in a hundred of naming the same list of brands. Claude repeated a list barely more often. The study is now more than six months old, but the 2026 measurements below show the same instability. Rand Fishkin notes that his coauthor, Patrick O'Donnell, works for Gumshoe.ai, an AI-tracking tool.

One measure held up despite this variation: the share of answers in which a brand appears. On a question about e-commerce marketing agencies, Google's AI named the agency Smartsites in 85 answers out of 95.

The answer also depends on who asks. In September 2026, Friction AI, a startup that measures brands in AI answers, compared anonymous accounts in the UK with accounts used for two weeks with detailed profiles, so the AI had learned about their users. ChatGPT and Gemini changed about one brand in five depending on what they knew of the user. Claude and Perplexity barely changed.

And each engine picks its own sources. In Rankability's August 2026 study, no two of the 11 AI engines had more than 24.1% of their cited pages in common.

Grow and Convert turns that into a method. Group your questions by topic, ask each one in several phrasings, run each many times, and read the result per topic and per engine, never as one average.

For Level AI, the share of answers naming the brand, averaged over every topic and engine tracked, was 17%. Its main topic told another story. Grow and Convert asked ChatGPT six phrasings of "call center quality assurance tools" every week. Over all those weekly runs, ChatGPT named the brand in 44% of its answers. Over the last six weeks, it named it in 83% of its answers.

One of those phrasings was run once a week for 36 weeks. On it, Google's AI Overviews named Level AI 74% of the time, and ChatGPT 39% of the time, with the recent runs well above the first ones. So the engine, the topic and the period each changed the figure, and one average across all three would have hidden these differences.

You can measure this by hand or pay a tracker to run the questions for you.

By hand, it comes down to three habits:

1. Each week, ask your questions several times each, in each of their phrasings and in the same conditions every time, in a private browser window or logged out. Keep the full text of the answers.
2. Note the share of answers that name you and the sites cited instead of you. Those sites are your contact list: that's where you need to be named.
3. Judge over several weeks, never on one answer.

Reading what the AI does behind an answer can also become a habit, beyond the counts. I try to read each answer from ChatGPT, Claude or Google, asking myself why it puts forward one page or brand rather than another. I sometimes ask ChatGPT a question just to see what comes out. That is how I try to understand how these AIs choose what they show.

Two reports complete the count by hand, both free as checked on 29 October 2026:

• Search Console's generative AI report. Announced on 3 June 2026 and open to all sites since 31 August, it counts impressions, the number of times your pages were shown in AI Overviews and AI Mode, by page, country, device and day. It gives neither clicks nor searches, and doesn't appear until your site has enough appearances.
• The AI Performance report of Bing Webmaster Tools. Bing Webmaster Tools is Microsoft's equivalent of Search Console. Its AI Performance report has been in public preview, an early version open to all, since February 2026. It counts your pages' citations in Copilot, in Bing's AI summaries and at some partners Microsoft doesn't name, and shows a sample of the searches the AI ran before citing you.

These are the two reports I use for the sites I run. In Bing Webmaster Tools, I look at the AI citations first, more than at the classic search figures. Bing Webmaster Tools is worth opening regularly for those AI answers: a lot of people use Copilot, and Microsoft's documentation says Copilot sends its web searches to Bing. Its report can show citations you didn't expect. Ahrefs shows me where my brands are mentioned. Its Web Analytics module also counts the visitors who arrived from an AI answer.

Then come visits and sales, starting with AI traffic, the visits an AI answer sends you. When someone clicks a link in ChatGPT, Perplexity, Gemini, Copilot or Claude, your analytics tool, the service that counts your site's visits, shows the visit as coming from chatgpt.com or perplexity.ai, for example. Visits from AI crawlers don't count as AI traffic: no person sees the page.

ChatGPT also adds a label to the end of each link it sends, utm_source=chatgpt.com, which analytics tools read, according to OpenAI's help for publishers. Your analytics tool can then show those visits apart: PostHog, an analytics tool, does, for example.

Two kinds of visits escape that count. Clicks from Google's AI Overviews and AI Mode are counted with Google's ordinary search clicks in Search Console, Google says, and its AI report gives no clicks. And some AI visits, from phone apps above all, arrive without saying where they came from and get counted as people who typed your address, Ahrefs noted in May 2026.

So the surest check is to ask new customers how they heard of you. Counted visits still matter: Charlie Whitworth's agency judges its GEO work on the AI visits it can count, plus the revenue they bring: "I can't really go to a client with a bunch of impressions", he said. If you want a paid tracker, this site's comparison of twelve AI visibility tools details what each one measures and what it costs.

What doesn't work, or stopped working

GEO recipes age fast. Several still sold today stopped working in 2026, or never had proof behind them.

Putting yourself first in your own "best of" list. For months, a common trick was to publish a "best providers" list on your own site with yourself in first place. In April 2026, Lily Ray said such lists could still sway AI answers, but that they were losing their place in search.

She had already seen "probably about 20 or 30 sites" with many such lists lose "a ton of traction" over the previous three months, and she expects search engines to keep demoting them. Since AI answers draw largely on search results, a list that loses its rankings tends to lose citations too. In the US, she adds, the FTC, the consumer-protection agency, expects you to disclose that you're ranking yourself.

Counting on Reddit. Reddit was long among ChatGPT's most cited sources. Promptwatch tracks the links ChatGPT cites. Reddit's share of those links was 3.83% between 18 July and 7 August 2026, and 0.52% between 14 and 17 August. That drop of 86% rests on four days ((3.83 − 0.52) ÷ 3.83 ≈ 86%). Promptwatch asked readers in August to treat it as provisional, and its page showed no update by 1 November 2026.

Posting as an ordinary user to recommend your own product also breaks Reddit's own rules, which ask for "authentic" participation, and Google counts such posts as inauthentic mentions.

Rewriting your pages for AI. A figure appears on many GEO pages: adding statistics, quotes and references would raise visibility "by up to 40%". It comes from a real study by Pranjal Aggarwal and colleagues at Princeton and other universities, posted in November 2023 and presented in 2024 at KDD, a data-science conference.

But those 40% come from a lab test. For each question, the researchers gave an AI the first five Google results, after rewriting one of those pages. The AI wrote its answer and showed, sentence by sentence, which pages it drew on. The researchers counted the words of the sentences that cited the rewritten page: the higher that count, the more the page weighed in the answer.

They compared that count with the one before the rewrite. A page could gain "visibility" this way by taking up more of the answer, without being recommended or placed higher.

In 2025, another team, of German and Korean AI researchers, tested the same idea again, in a different way, with four more recent AI models. It presented its results at NeurIPS, the main AI research conference.

This time the researchers didn't count words. They gave the AI a question and a list of pages, and the AI answered by citing some of those pages, one after another. The researchers noted where the rewritten page came among those sources. If it came earlier than the original, unrewritten page, the rewrite had worked.

They tried several ways of rewriting a page, such as adding statistics or references, or making the tone more authoritative. They tested each way on six types of subject, from news to books.

One way of rewriting, tested on one type of subject, counts as one case: the researchers measured 54 cases in all. Only 3 of the 54 cases showed a gain too large for chance. In all the others, the rewritten page came no earlier among the sources, or so little earlier that chance could explain it.

Adding statistics even pushed the page back. The researchers crossed their four AI models with their six types of subject, 24 combinations in all. In 19 of them, the page with statistics was cited later than the original page.

The researchers also tried rewriting nothing: they only put the page first in the list handed to the AI. That is the place a page gets when it ranks first for the search the AI runs. With that move alone, the AI cited the page earlier than with any rewrite: in the lab, ranking beat rewriting.

Jan-Willem Bobbink, an SEO consultant, ran his own test. Louise Linehan, who carries out studies for Ahrefs, reported it on the Ahrefs podcast in September 2026.

Jan-Willem Bobbink took two pages. He gave the first "the GEO treatment": statistics, an authoritative tone, references to sources. He left the second as it was. In the end, the untouched page got more citations. One test on two pages proves little, but it points the same way as the 2025 test presented at NeurIPS: the page rewritten for AI wasn't cited more.

These tests don't mean your own figures are useless. The researchers and Jan-Willem Bobbink added statistics to a page to impress the AI. Your own data is something else: your prices, your tests and your customer case studies are information the AI can't find anywhere else.

Counting on an llms.txt file. An llms.txt file is a plain-text summary of a site, published at the address /llms.txt for AI tools. Jeremy Howard, cofounder of the AI lab Answer.AI, proposed it in September 2024. Coding agents, the AIs that write software, use such files to read the documentation of the tools they work with. For showing up in Google, AI Overviews included, the file doesn't help: Google writes in its guide that Google Search "doesn't use" such files.

Ahrefs looked at the server logs of 137,210 sites in May 2026: 28%, about 38,000 sites, published an llms.txt file. 97% of those files were never requested, by a bot or a person.

I still publish an llms.txt and an llms-full.txt, the version that holds the text of every page, on all the sites I run. I have seen Claude go and read one of these files to answer a question I asked it. An llms.txt file costs little to publish, but it replaces no other work.

Adding special markup for AI. Structured data (schema markup) helps your pages earn rich results in Google, but Google writes that its AI features need "no special schema.org markup". In Rankability's study, no type of structured data improved a page's place among the citations.

Mass-producing pages with AI. In April 2026, Lily Ray said she sees sites that publish large volumes of AI-written pages rise in Google for three to six months, then decline "like clockwork", with their AI citations falling after them.

How much AI traffic there is, and what GEO is worth

The method aims at citations and mentions. What they bring comes in two forms: visits from AI answers, and buyers who read your name and come to you later.

The visits are still few. In September 2026, Ahrefs measured the visits to the 115,254 sites that use its free Web Analytics tool. AI assistants sent 0.43% of those visits, and ChatGPT alone 0.34%. Search engines sent 26.82%, and Google alone 23.96%. So Google sent about 56 times as many visits as all the AI assistants together (23.96 ÷ 0.43).

These are sites whose owners chose an SEO company's analytics tool, a sample that leans toward people who care about search.

SE Ranking, another SEO software company, found figures on the same scale on 101,574 sites from January to April 2026: 0.32% of their visits came from AI platforms worldwide, up from 0.02% in 2024, and 0.29% in the US. ChatGPT sent about three AI visits in four.

Kevin Indig put AI traffic in July 2026 at "maybe 3% of your total organic traffic if you're lucky". On the sites Ahrefs measured, AI visits equal about 1.6% of the visits from search engines (0.43 ÷ 26.82). He considers traffic the wrong yardstick: people rarely click a citation, he explains, they open a new tab and go straight to the site. One of his clients gets 10% of its new customers from AI.

The value lies elsewhere: a buyer reads your name in an answer, then searches for it or comes straight to your site. Citations alone don't prove that value, though. Will Reynolds, founder of Seer Interactive, a US search agency, got a page cited 19 times more often by AIs and saw its traffic only double. He judged the value to his business negligible, as Louise Linehan reported in September 2026. She didn't give the starting figures.

Being named in the answer matters more because the click on the first result is shrinking. In data from December 2025, published in February 2026, Ahrefs compared 300,000 keywords, the searches people type, before and after AI Overviews.

On searches without an Overview, the first result's click rate fell, the share of searchers who clicked it, from 7.6% in December 2023 to 3.9% in December 2025. On searches that now show one, it stood at 7.3% in December 2023. Had those searches followed the same trend, their first result would have kept a click rate of about 3.7% (7.3% × 3.9 ÷ 7.6). It got 1.6%.

Ahrefs, working from unrounded data, puts the gap at 58%: "For every 100 clicks you could historically earn for a top-ranking page, Google now 'keeps' 58." It's a correlation, measured by a company that sells SEO software. Seer Interactive reported drops of 49% to 65% in organic click rates in a September 2025 update cited by Ahrefs.

Are visitors from AI better customers? The studies cited here, all from 2025 and so more than six months old, disagree:

• For. At a client of Seer Interactive, 15.9% of visitors from ChatGPT completed one of the goals the site tracks, against 1.76% of visitors from Google. Patrick Stox of Ahrefs reported in June 2025 that on Ahrefs' own site, visitors from AI search were 0.5% of the traffic but brought 12.1% of sign-ups. By Ahrefs' own calculation, they signed up 23 times as often as visitors from ordinary search.
• Against. A study published by Search Engine Land, a trade publication on search marketing, in October 2025, on 973 online stores, found that visitors from ChatGPT and other AI assistants bought less often than visitors from Google Search, and brought less revenue per visit.

Joe Lazer, part-time marketing director at the content marketing company Pepper, offers an explanation: the favorable studies cover companies that sell to other businesses, the unfavorable one online shops. For software or a professional service, the AI often sends someone who has already compared offers to the seller's site. For a consumer product, it tends to send them to review sites, so shoppers who reach a store from an AI are often still comparing.

On my own sites, what I've seen matches the favorable studies. AIs have cited my pages and customers have come through their answers: I saw it in the link they arrived by. In my case, those customers buy more readily than the others, I think because they trust their AI when it recommends a service. That's my observation, not a rule.

Your action plan

Every action in this guide, in the order of the steps, to tick as you go.