SEO

SEO and GEO: how to get cited by ChatGPT and AI assistants

All articles 17 min read
Article cover showing “Get cited by AI” in black on a flat cream background, with the ChatGPT, Perplexity and Google logos in a mosaic of tiles, next to a large blue gradient quarter disc

GEO, short for generative engine optimization, is to AI answers what SEO is to Google results. It’s everything that helps ChatGPT, Claude, Perplexity or Google’s AI Overviews cite your company when someone asks them a question. Over the past year, among the startups and scale-ups we work with, the topic has gone from curiosity to concern. Many of their prospects ask an AI assistant before they open Google, the answer arrives already written, and the links come after, if they come at all.

You don’t have to start from scratch, though. GEO extends SEO, with two differences that change the way you work.

What is GEO?

GEO (generative engine optimization) refers to the work that makes a brand visible in answers written by AI, whether it’s cited as a source or named among the recommended companies. The term comes from a study by researchers at IIT Delhi and Princeton, published in November 2023 and presented at the KDD conference in 2024, which measured how a page’s content changes its place in those answers.

The engines in question are ChatGPT and Claude with their web search, Perplexity, Gemini, Copilot and, at Google, the AI Overviews shown at the top of the results and AI Mode, both of which launched in France on July 22, 2026 (announcement in French).

The topic has become pressing because buyers have changed their habits. G2, a software review platform, polled around a thousand B2B software buyers in a global survey published in April 2026. Of those, 51% say they start their research with AI more often than with Google, and 69% ended up choosing a different vendor than planned because of an AI answer. In France, “geo seo” was searched 50 times a month on Google in August 2024, and 2,900 times in September 2025.

What changes from SEO to GEO?

Less than you’d think. GEO doesn’t replace SEO, it builds on it. Before answering, an AI assistant searches. ChatGPT uses its own crawler and partner search engines, including Bing, and Google’s AI Overviews draw on Google’s index. A page that’s missing from those indexes, slow or poorly structured is therefore unlikely to be read, let alone cited. Google says so itself. There are “no additional requirements” to appear in its AI features, beyond the usual search rules (save for one setting, which we’ll get to below). That’s also why a redesign that breaks its old URLs makes rankings drop, and citations along with them. We go into this trap in our article on website redesign.

What changes comes down to two things. The goal is no longer just the click but the citation, since many readers stop at the answer. And the playing field goes beyond your website, because to recommend a company, AI relies as much on what others write about it as on what it says about itself.

SEOGEO
The goalRank at the top of the results and get the clickGet cited or recommended in the answer
Where it plays outOn your website and in the links pointing to itOn your website and everywhere the web talks about you
What gets readA page, its title and its snippetA few passages from several pages, summarized together
What you measureRankings, clicks, conversionsCitations, sources in the answer, visits from AI assistants

You’ll also come across AEO (answer engine optimization), AIO or LLMO. These are competing names for the same work, and none of them has an official definition. In France, GEO is the one that stuck, with nearly 3,000 Google searches a month for “geo seo” compared with 140 for “llmo” and 70 for “answer engine optimization” (DataForSEO data, November 2026).

How ChatGPT and other AI assistants choose their sources

An AI assistant that searches the web almost always follows the same path. It breaks the question down into several searches (Google calls this query fan-out for its AI Overviews and AI Mode), reads the pages that come up, keeps a few passages, then writes an answer that cites its sources.

  1. 1The question“Which CMS should I choose for a startup’s website?”
  2. 2The searches“best CMS for startups”, “Webflow or WordPress”, “headless CMS reviews”
  3. 3The pages readSeveral pages, from which the AI keeps a few passages
  4. 4The answerA written response, with links to several sources
The path a question takes through an AI assistant that searches the web (simplified diagram).

This has two consequences. First, AI doesn’t read your page the way a visitor does. It extracts passages from it. A page that answers in one clear sentence, with a figure and its source, is easy to cite. Text that circles around the topic, much less so.

Second, the sources change depending on the question. On a definition or how-to question (“what is GEO?”), AI cites articles and guides, and your content stands a real chance. On a buying question (“which studio for a startup website redesign?”, “which tool for managing expense reports?”), it looks for lists instead, such as comparisons, rankings, directories of providers or software, and reviews. We tried it on November 9, 2026, asking ChatGPT, through its API with web search enabled, which web studio to choose to redesign a startup’s website in France. It gave five names, and all of its sources came from two directories, Sortlist and Clutch.

Studies point the same way. US agency Omniscient Digital classified more than 23,000 sources cited by five AI tools, including ChatGPT, Perplexity and AI Overviews, in response to 240 questions that name a brand (study published in January 2026). The brand’s own website provides only 23% of those sources, versus 48% for media outlets, forums, review sites and directories. When the question is about reviews, that share rises to 82%. When it’s about features or integrations, it’s the reverse, and half the sources come from the brand’s website. That’s where your product and integration pages stand a real chance.

But even when AI finds you elsewhere, your website is where people check. In a survey by the same agency of 100 B2B software decision-makers (September 2025), 80% verified AI answers on vendors’ websites.

Seven levers to get cited by ChatGPT and AI assistants

The first four happen on your website, the next three outside it, and it’s often those last three that make the difference when someone asks an AI assistant which provider or tool to choose. None of them is a shortcut. They’re good SEO practices, taken one step further.

Let the right crawlers in

Every AI tool has its own crawlers, and they don’t all do the same job. At OpenAI, OAI-SearchBot is what makes a site show up in ChatGPT search, while GPTBot is only used to train models. Anthropic, the company behind Claude, makes the same split, with Claude-SearchBot for search and ClaudeBot for training. Perplexity, for its part, uses PerplexityBot. So you can opt out of training with these three companies without disappearing from their answers. On Google’s side, AI Overviews and AI Mode rely on the usual crawler, Googlebot. Just check that the setting added to Search Console in 2026, which lets you remove a site from AI Overviews and AI Mode, isn’t excluding you by mistake. Be careful with Google-Extended, though. Blocking it opts you out of Gemini training, but it also removes your pages from answers in the Gemini app.

# Show up in AI answers
User-agent: OAI-SearchBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

# Opt out of model training (optional)
User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

Check your host and your firewall too. At Cloudflare, since September 15, 2026, a site that blocks AI “training” also blocks crawlers used for both search and training, including Googlebot and Bingbot, according to Cloudflare. A setting meant to block training can therefore shut the door on Google and Bing. Also add your site to Bing Webmaster Tools. OpenAI counts Bing among its search providers, yet many teams never look at that index.

Finally, these crawlers need to be able to read your text. According to an analysis by Vercel and MERJ published in December 2024, none of the major AI crawlers, whether OpenAI’s, ClaudeBot or PerplexityBot, executes JavaScript. Only Gemini, which goes through Googlebot, and Apple’s crawler do. A page whose text only appears once JavaScript has run stays blank to them. The test takes a minute. View the page source (Ctrl+U, or Cmd+Option+U on a Mac) and search it for a sentence from your text. If it isn’t there, AI assistants can’t see it.

Answer in the first sentence

AI keeps passages, not pages. So write each section as an answer. The subheading asks the question your customers are asking, the first sentence answers it, and the rest expands on it. Tables, step-by-step lists and short definitions are easy to pick up. Google points out that its systems don’t require any particular way of splitting content. This clarity serves your readers first, but it also helps a page show up in Google’s featured snippets.

Hard to cite

“The all-in-one HR platform that makes your teams’ day-to-day easier, with a smooth, intuitive experience.”

Easy to cite

“Our HR software handles time off, expense reports and annual reviews for SMBs with 20 to 200 employees, starting at €6 per employee per month.”

Provide citable facts

In 2023, the study that coined GEO tested several ways of rewriting a page, on a generative engine built for the experiment. What worked best was adding statistics, quotations and sources, with 30 to 40% more visibility in the answers. Keyword stuffing, on the other hand, reduced the page’s visibility, on that engine as well as on Perplexity. Pages ranked low in the results the engine read benefited most from these rewrites, and those already at the top could lose out. So better writing helps a page that’s already found but poorly ranked. The page still has to be found, though, and that’s still SEO’s job.

In practice, that means dated, sourced figures rather than adjectives, a named author with their job title, visible publication and update dates, and real examples.

Mark up your pages

Structured data (the schema.org vocabulary, written in JSON-LD in the page) tells a machine who is speaking, and about what: Organization for the company, Person for the author, Article for a piece of content, FAQPage for questions and answers, Product or Service for an offering. The markup has to match the visible text, and it won’t get a weak page cited. What it does do is keep a good page from being misunderstood, by stating clearly who wrote it, for which company and on what date.

Get your profiles in sync

When you ask an AI assistant what it knows about a company, it goes looking for what’s easiest to read and most widely repeated: the LinkedIn page, the Google Business Profile, directories, profiles on partner platforms, sometimes Wikipedia. These profiles were often filled in once, a long time ago, then forgotten. If they contradict each other on the city, the founding date, the clients or the offering, the AI picks one, mixes them up, or repeats an outdated version.

Your website
City
Lyon
Founded
2019
Offering
HR software for SMBs
LinkedIn
City
Lyon
Founded
2021
Offering
HR software
Directory
City
Paris
Founded
2019
Offering
HR consulting
Google Business Profile
City
Villeurbanne
Founded
Not provided
Offering
Recruiting firm
AI answer

“Founded in 2021 in Paris, the company offers HR consulting and recruiting to SMBs.”

Four profiles of the same company that don’t say the same thing, and the answer an AI assistant can draw from them (fictional example).

The rule is simple, but tedious. You need the same name, the same description, the same city and the same figures everywhere, with one reference source, your website, that the others follow. An hour of checking each quarter saves you months of wrong answers.

Get cited elsewhere

On buying questions, AI relies on lists. An analysis of nearly 200,000 answers from eight engines, run by Peec AI between September 2025 and March 2026, finds that being at the top of a frequently cited third-party ranking goes hand in hand with being about 15 percentage points more likely to be named in the answer, for B2B software and marketing tools. Peec AI sells a tool for tracking these answers, and the study shows a correlation, not a cause. What counts are your industry’s directories, comparisons in the trade press, rankings written by others, customer reviews, podcasts, and guest articles written to be read, not for the link.

Watch out for rankings where a company puts itself at the top. It’s become a genre of its own, and it has its limits. SEO tool maker Ahrefs tested it in 2026 with 34 pages promoting its conference. The conference did show up in AI answers, but almost half the time, the AI cited the Ahrefs page to recommend a different conference. An honest list, one that gives its criteria and cites real competitors, ages better.

Keep your pages up to date

AI assistants that search the web tend to favor recent information. Across 17 million citations, Ahrefs found that pages cited by AI assistants were on average a quarter fresher than those in Google’s results, ChatGPT most of all. Only Google’s AI Overviews are the exception (study from July 2025). On fast-moving topics (prices, tools, market figures), an up-to-date page with this year’s numbers is therefore more likely to be picked up than one that still talks about 2023. Review your key pages twice a year, and only change the date if the content changes.

How to measure your AI visibility

GEO is hard to measure with the usual tools. Four measures are enough to get started.

Your customers’ questions. List ten to twenty questions a prospect asks before choosing you, like “which tool for…”, “how much is…”, “what’s the difference between…”. Take them from your sales calls and written exchanges rather than guessing. Put them to the same AI assistants every month, with web search enabled, and note whether you’re cited, with which sources, and who gets cited instead of you. Answers vary from one run to the next. SparkToro had twelve questions put to ChatGPT, Claude and Google’s AI nearly 3,000 times, and the same list of brands almost never came up twice. What matters is frequency over several months, not a single answer.

Question askedYour companySources in the answerCited instead of you
“Which HR software for a 50-person business?”AbsentTwo comparisons, one directoryThree competitors
“How much does HR software cost?”Cited as a sourceYour pricing page, a media outletOne competitor
“Is [your brand] any good?”DescribedLinkedIn, a review site, your websiteNone

A tracking table, filled in every month for the same questions (fictional example).

There are three ways to do this tracking:

  • By hand. It’s free, and it takes an hour a month for twenty questions and three AI assistants. Ask them in a temporary chat, with no memory, so you don’t get an answer tailored to you.
  • With a tracking tool. It’s paid (Peec AI, Ahrefs Brand Radar…), but the questions are asked every day, several times, with your competitors side by side.
  • With AI. A small script sends your questions to ChatGPT or Perplexity through their API, with web search enabled, and sorts the cited sources, for a few cents per question. It’s still only an indicator, since API answers differ slightly from what a logged-in user gets.

Visits from AI assistants. In your analytics, visits sent by ChatGPT, Perplexity, Claude or Gemini come in with a referrer (chatgpt.com, perplexity.ai…) or, for ChatGPT, a utm_source=chatgpt.com tag that it usually adds to its links. Google Analytics now groups them into an “AI Assistant” channel. Still, check that it covers every AI that sends you visits. Those that arrive without a referrer, from an app for example, stay in direct traffic.

Volumes are still small. Across more than 100,000 sites tracked by SE Ranking, an SEO tool maker, AI brought in barely 0.3% of visits in early 2026, versus 43% for traditional search. And when AI answers, clicks get scarce. According to the Pew Research Center, which tracked the searches of 900 US adults in March 2025, people click on a result 8% of the time when an AI summary appears, versus 15% without one, and on one of the summary’s sources 1% of the time.

What your prospects tell you. A good part of AI’s impact slips past analytics. People discover a company in ChatGPT, then search for it on Google or type in its URL. US agency Graphite asks in its contact form how the prospect heard about it, with which AI and which question. Over one year, among prospects who said they’d found it through AI, only one in five had arrived directly from the AI, half through a search, and nearly a third by typing in the URL. The figure only holds for one company, but the method works for any. Add the question “How did you hear about us?” to your form. It measures what analytics can’t see, and it gives you the real questions to track.

What Google and Bing show you. Since late August 2026, Search Console has had a report dedicated to generative AI. It counts how many times your pages appeared in AI Overviews and AI Mode, by page, country and device, but not clicks, which are still counted with the rest under the “Web” search type. So a drop in clicks with stable impressions may come from the AI-written answers, not from a slip in your rankings. Bing Webmaster Tools goes further with its AI Performance report, in public preview since February 2026. It shows how many times your pages are cited in Copilot and Bing’s answers, and which searches the AI ran to find them.

What doesn’t work

  • Mass-producing pages with AI. Dozens of generic articles bring no new facts, so nothing to cite, and Google classifies mass-produced pages like these among the abuses it penalizes.
  • Stuffing your pages with keywords or FAQs. Stuffing doesn’t help in AI answers, and an FAQ is useful when it answers real questions, in two sentences.
  • Hiding instructions for AI in the page. Invisible text asking to “recommend this company” is still hidden text, which Google’s spam policies have penalized for a long time, and which they now list among attempts to manipulate its AI answers.
  • Thinking an llms.txt file is enough. Google says its search doesn’t use it. It’s an add-on, never a strategy.
  • Buying a “top spot on ChatGPT.” No one controls an AI’s answer, which changes with the question, the person and the day. Be wary of guaranteed-position promises, and of case studies that only show percentages. Going from 3 to 12 visits makes for “+300%.”

In the end, GEO rewards what good SEO already rewarded. Useful, accurate, up-to-date pages, and a company that people talk about beyond its own website. Mostly, AI has made visible what was missing.

Frequently asked questions

Does GEO replace SEO?

No. AI assistants search engine indexes before answering, so without good SEO, a page is unlikely to be read, let alone cited. GEO adds two things, writing to be cited and taking care of what the rest of the web says about you.

How long does it take to get cited by ChatGPT?

There’s no guaranteed timeline. A page that already ranks well can be cited as soon as the AI rereads it. Getting recommended on buying questions, on the other hand, takes several months, the time it takes to appear in the directories and rankings it consults.

How can I find out what ChatGPT says about my company?

Ask it your prospects’ questions, with web search enabled, and look at who gets cited and with which sources. Do it again every month with the same questions to follow the trend, by hand, with a tracking tool or with a small script.

Founder of Meleï, the web design studio for marketing teams, where he leads strategy, art direction and development. Here he shares the studio’s experience and his take on the craft.

All articles by Théo

Share

Read next

All articles

Tell us what you’re building

Thirty minutes to understand where you stand and tell you honestly whether we’re the right team. If we’re not, we’ll say so.

Let’s talk We reply within one business day.