ChatGPT SEO: Getting Cited by ChatGPT, Perplexity, and Gemini, What Actually Decides It

You type your business into ChatGPT, with your city, the way a prospect would. The answer comes back in six lines, clean and confident. It cites three companies. None of them is yours.

The next reflex is to look for how to do "ChatGPT SEO," the way you'd look for how to move from position 8 to position 3. Except the question is framed wrong: there's no ranking inside ChatGPT. No SERP to climb. There's a selection, redone with every answer. Two people asking the same question the same day can get two different answers, with two different sets of sources. Ask again later and the text may have shifted.

You're no longer reasoning in terms of position, but in terms of the probability of being picked up. In this article, we look at the mechanism: what happens between the question and the answer that gets shown.

Absent, Mentioned, Recommended: Three States, Not a Score

You need a vocabulary for this: "being visible in AI" means nothing until you say to what degree. This is the scale that matters:

• Absent: neither your name nor your pages appear in the answer.
• Mentioned: your name appears, but the source shown isn't you (a directory, a comparison site, a press article).
• Recommended: the engine puts you forward as the answer, not just as a name mentioned in passing.

"Being cited" isn't added as a fourth state: it describes the mechanism. One of your pages appears among the sources shown, with the link. You can be cited without being recommended, and the reverse.

The case we run into most often, and the most misunderstood, is "mentioned without the source being you." A lot of people see their name go by and think they're doing well: it's other people's sources carrying it, and those are what the engine will keep reading. The progression runs in this order: Absent to Mentioned to Recommended. You don't skip a step.

What Happens Between Your Question and the Answer

No one outside these systems knows their internals. The visible chain, though, can be observed. Four steps.

• The engine rephrases. Your sentence isn't the query that gets run: "I'm looking for an accounting firm in Lyon, ideally one that specializes in startups" becomes one or several short searches, worded in the engine's own way.
• It retrieves a set of sources. How many, you can't see: only the sources shown are observable, and nothing says they're all the ones it read.
• It keeps certain passages. Not whole pages: chunks that answer the question.
• It writes and attributes, based on those passages, with the sources shown when it did search.

The consequence is blunt: the question you think you're targeting isn't the one the engine actually runs. Optimizing word for word around the prospect's exact phrasing does nothing; covering the topic so you land in the net regardless of the rephrasing does.

Two Filters, Not One

To be cited, you have to clear two successive filters, and they aren't worked on the same way.

Filter 1: retrieval. Your page has to be reachable by the engine's agents, exist in an index, and match the rephrased query. If you fail here, nothing else matters.

Filter 2: reuse. Among the retrieved sources, the engine only keeps passages that can be reused as is: a paragraph that answers the question, makes sense out of context, and doesn't force the reader through three screens before getting to the information.

Clearing filter 1 without filter 2 means being read and discarded. The typical case: the service page that talks about itself, "since 2011, our passionate team has supported you with a tailored approach." The engine retrieved it, found no quotable sentence in it, and took the one from the comparison site next door instead, which gave a price and a timeline.

What you usually read online only addresses one of the two filters: the technical advice stops at the first, the editorial advice already assumes it's been cleared. That's why, applied on their own, each seems to produce nothing.

The Three Engines Don't Decide the Same Way

Same mechanics, different trade-offs. We're sticking to what the providers publish and what can be observed by asking the questions.

ChatGPT: The Engine That Doesn't Always Search

ChatGPT can answer without going to the web, purely from what it learned: no sources, no links, and your only chance of showing up is if your name is already associated with your topic. There's no direct short-term lever there.

When it does search, it shows its sources and becomes addressable: both filters apply, and your pages can get in. The same prompt doesn't always trigger the same mode. When you test, note whether sources were shown, otherwise you're comparing two different things.

Perplexity: Search First

Perplexity is built around search: numbered sources are close to the default. It's the most legible engine to diagnose with, even if your market lies elsewhere. What it cites tells you which pages cleared its own two filters, whether yours or someone else's. Don't draw any conclusion from that about ChatGPT or Gemini. Each has its own index. Perplexity shows you the mechanism, not other engines' verdicts.

Gemini: The Google Anchor

Don't conflate two surfaces: the Gemini app on one side, and the generated answers inside Google Search (AI Overviews and AI Mode) on the other. For the latter, Google documents that there's no dedicated bot and no special markup to set up: it's the search index that's used, your page has to be in it and stay eligible for a snippet, and the control remains Googlebot's robots.txt (Google Search Central, AI features, accessed August 6, 2026). What decides which sources get shown isn't documented. Test the two surfaces separately.

What the three have in common is more useful than what separates them: they cite what they can reach, and they favor passages that answer the question. A good share of the sources shown are third parties rather than the brand's own site. That's our observation, not a market-wide measurement. If you sell in France, add Mistral's Le Chat to your tests: it's a player in the French market.

The Five Levers That Actually Decide It

For each one: what it is, how you check it on your own site, and which filter it acts on.

Access

Filter 1, disqualifying. An engine won't cite a page it can't read. Check your robots.txt, your firewall, your CDN, and any leftover blocks from a vendor who "secured" the site two years ago.

One point almost nobody separates out, even though providers document it clearly: the bots that crawl for training and the agents that fetch your page when a user asks a question are different. Per OpenAI, Perplexity, and Google documentation, accessed August 6, 2026:

• OpenAI publishes four agents: GPTBot for training, OAI SearchBot for the indexing that feeds ChatGPT search, ChatGPT User for visits triggered by a user, and OAI AdsBot for advertising pages.
• Perplexity publishes two: PerplexityBot, which indexes for search and which Perplexity states isn't used to train models, and Perplexity User, triggered by a user.
• Google doesn't publish a dedicated agent for its AI answers in search: control runs through Googlebot. Its Google-Extended token covers two things at once: training future Gemini models and grounding your pages in Gemini apps. Google states it changes neither inclusion in search nor ranking.

Blocking the training crawlers is legitimate; blocking the agents triggered by a user's question makes you uncitable. Google-Extended is the trap in this rule: it doesn't separate the two. Turning it off to refuse training also removes you from Gemini grounding, and so from a citation surface.

Second trap: robots.txt doesn't control everything. Because the visit originates from a user, OpenAI states its rules may not apply to ChatGPT User, and Perplexity states that Perplexity User generally ignores them. For those, what decides is your firewall and your CDN. These names and rules change: check them directly with each provider when you act, not in a blog post, including this one.

Is the Answer Actually on the Page, in Plain Text

Filter 2. At the top of every important page, write a self-contained passage: two or three sentences that answer the question before you expand on it, understandable out of context. Word your headings like your customers' questions, not like your offer's product names.

Spell out the facts in the actual text of the page: prices, terms, timelines, scope, what you don't do. A figure locked inside a PDF isn't invisible. Google documents that it indexes the text of such files (Google Search Central, indexable file types, accessed August 6, 2026). But then the file becomes the source, not your page. Write the figure in plain text in the HTML. Simple test: paste a single paragraph into a document on its own. If it means nothing on its own, it won't get picked up.

Consensus Beyond Your Own Site

Both filters at once. Answers cite third parties: comparison sites, directories, trade press, forums, reviews. If your brand exists nowhere but on your own site, you can be flawless at home and still be absent.

How to check: ask your questions, note who's cited in your place. Not necessarily your competitors, mostly sources. That's your roadmap. Trusted reference platforms carry particular weight here: they are the consensus sources that AI engines reach for when they need a neutral, verifiable anchor.

Reddit is the sharpest example of consensus you can't fake from your own site. On comparison and "is it worth it" questions, a Reddit thread is very often among the sources shown: engines treat it as unvarnished experience, the one register a service page can't imitate. Showing up there means disclosed participation, answers that respect each subreddit's rules, and a history in the threads your market reads; done that way, Reddit keeps feeding the sources the engines consult long after the thread itself has cooled. This is a discipline of its own, which is why specialist partners such as Red Visibility, America's #1 Reddit Agency build Reddit visibility for brands that need the consensus without tripping the communities.

Freshness, and Whether It's Legible

Filter 2. A page can be recent without showing it. Display a publication date and an update date that match content that's actually been refreshed: changing the date without changing the text is just noise. As far as we know, no provider publishes a freshness criterion. So no one can sell you a numeric rule about it. What holds true without a measurement: on anything that moves (prices, regulation, comparisons, "as of 2026"), an undated page tells no one whether it's still valid.

How to check: on your strategic pages, compare the date shown against the date of the last real content change.

Entity Consistency

Filter 2, indirectly, through other people's sources. Same name, same activity, same address, same figures everywhere: your site, your business listing, your LinkedIn, directories, comparison sites. We haven't measured the effect of a contradiction on a citation, and we won't invent one. What we can say: when two sources say two different things about you, there's no single version to pick up.

How to check: search your company's exact name, open the first five listings that mention you without being you, and compare them against your "about" page.

What's Being Sold to You That Proves Nothing

The GEO market is moving faster than the evidence for it. A few claims to receive with caution.

The llms.txt file. As of today (August 2026), no engine provider documents using it to decide on a citation, and Google has weighed in on this: John Mueller, on Bluesky on June 17, 2025, said "FWIW no AI system currently uses llms.txt." Setting it up takes a few minutes and breaks nothing. Billing you for it while promising citations is selling a hypothesis at the price of a result.

Schema.org markup sold as the golden key. Structured data does condition certain rich results in classic search: that's basic hygiene. But for its AI answers, Google states that no additional technical requirement is added on top of normal search eligibility (same documentation). Do it, but don't build your strategy on it.

Next-generation keyword stuffing, the kind that repeats "best [profession] 2026" across every H2: that's optimizing for a ranking that doesn't exist here.

Buying "AI mentions." Advertising inside answers does exist, and it's openly stated: Google announced on May 20, 2026 two ad formats built into AI Mode, labeled as sponsored and in testing as of that date; OpenAI documents a control agent for advertising pages on ChatGPT. That can be bought as advertising, and it looks like advertising. The rest of what's sold under this name is placement within third-party sources. Call it what it is.

The guarantee of being "first in ChatGPT." Impossible by construction: there is no first, and the answer varies from one run to the next.

In the same vein, counting prompt volumes doesn't substitute for actually reading the answers: what's said in the sentence where your name appears matters more than how many times it appears.

One last and most important point: the criteria for selecting sources aren't published. Everything written above comes from provider documentation or from repeated observation of the answers. Far from being an admission of weakness, saying so is what separates an observation from a belief.

Checking Where You Stand, Engine by Engine

The principle is useless without a starting point. A method you can run yourself, no tool required.

Write down ten questions your customers actually ask. Not your keywords: their sentences, in their words. Ask them on each engine and note four things: your verdict (absent, mentioned, recommended), whether one of your pages appears among the sources shown, who's cited in your place, and whether the engine searched or answered from memory.

That third column is the valuable one: it tells you which pages your market treats as legitimate, and so where you need to exist.

A Reading That Holds Up: Always Note the Engine and the Date

An AI answer with no engine name and no timestamp counts as an anecdote, not as evidence. Note both, and repeat the measurement regularly: it's the only way to tell a trend apart from a fluctuation.

Where to Start This Week

Four actions, in this order.

• Run a baseline measurement, engine by engine, dated. Everything else will be judged against this line.
• Unblock access. robots.txt, firewall, CDN. If the agents can't get through, everything else is decoration.
• Rewrite your two most strategic pages: a citable passage at the top, facts spelled out in plain text, headings worded as questions.
• List the sources cited in your place and go establish a presence there, current and consistent.

Absent, then mentioned, then recommended. The first step can be measured today.

Questions to Unroll

This article covers the mechanism and the levers. Each of the questions below deserves its own deep dive, and we'll be unrolling them one by one on this blog.

• How do you measure AI visibility over time without turning prompt testing into noise?
• Which bot and agent rules should your robots.txt actually contain in 2026, provider by provider?
• How do you write a self-contained citable passage, and what does one look like on a real service page?
• Which third-party sources matter most in your market, and how do you earn a place in them?
• How do trusted reference platforms act as consensus anchors for AI engines, and what does it take to earn them?
• When does entity inconsistency actually cost you citations, and how do you fix it across the web?
• Should you advertise inside AI answers, or is sponsored placement in an answer a credibility risk?
• Le Chat and the French market: how does Mistral's assistant decide what to cite?

Frequently Asked Questions

Make Your Narrative Count

Narrative moves markets, and the engines that answer your prospects' questions now decide whose version of the story gets repeated. Elyseum works with leaders and companies to build the presence that gets them picked: access that's open to the right agents, pages that answer before they explain, and a consensus across the sources the engines trust. Contact our team for an assessment of where you stand, engine by engine, and a plan to move you from absent to mentioned to recommended.