How to get your brand mentioned in ChatGPT (2026)

The standard answer is that you should publish authoritative content and build a strong brand. That is not advice. It describes a company that is already winning, and it cannot be wrong, which is most of what is wrong with it.

It is also the one tactic that has been tested against a control and found close to worthless. We will get to the numbers.

There are two ways a brand ends up in an AI answer. It gets retrieved, when a page mentioning you is fetched while the answer is being written. Or it gets remembered, when the model names you without searching at all. Retrieval follows your rankings and responds in days. Memory follows what has been written about you for years, and responds in training cycles. Most wasted budget in this channel comes from working the wrong one.

The model searches, but it decides what to search for

Retrieved. The assistant rewrites the prompt into ordinary search queries. It runs them, reads the results, and summarises. You appear because a page naming you came back. This is the query fan-out. It behaves like SEO aimed at a moving target.

Remembered. The model answers without searching. You appear because it already associates you with the category. This is parametric visibility. It moves only when the model’s knowledge cutoff moves.

The two are not independent. The remembered half gates the retrieved half. The model writes its own search queries, and it reaches for brands it already knows. A brand it has never heard of never becomes a candidate. The search that would have found you never runs.

Step 1: Find out which half is broken

Ask each model the same buying question twice. Once through the API with browsing off. Once in the product with search on. Use the question a buyer would ask, not your brand name.

Search off	Search on	Diagnosis	Where the work is
Absent	Absent	Not known, not findable	Both halves. Retrieval first
Present	Absent	Known, but losing the retrieval race	Steps 2 to 5
Absent	Present	Findable, not established	Step 6
Present	Framed badly	Not a visibility problem	Sentiment, and the sources doing the framing

The fourth row matters more than people expect. Being named as the cheap option is not the same problem as being absent. More content will not fix it.

One caution. A single run tells you almost nothing. In the most disciplined public measurement we know of, sixteen frozen prompts run daily for twenty-eight days, the spread between runs was 2.8 percentage points. Repeat the test over several days before believing it.

Step 2: Find the queries the engine actually runs

You are not competing for the prompt. You are competing for the rewritten queries behind it. Those are not stable. They differ between engines, and they drift over time on the same engine.

This is the usual reason a brand is invisible for a question it should own. The engine never searched for the words the page was written for. It searched for a longer variant, or a comparison, or the same thing with a year on the end.

Get the fan-out queries for your important prompts. Then treat each variant as a keyword, because that is what it is. Ranking just outside the top ten means adding a section to the page that already ranks. A variant with genuinely different intent deserves its own page. A reworded version of your head term deserves nothing.

Step 3: Win the pages you do not own

This is the biggest lever. Most teams skip it, because it does not feel like marketing.

What an assistant retrieves for a category question is mostly not vendor websites. It is roundups, comparison articles, review directories and forum threads. Reddit appears constantly, not because models love forums, but because Reddit ranks in the index the engine is borrowing. Search Google for how to get mentioned in ChatGPT. The top result is a Reddit thread, above every agency guide on the subject.

So the question is not what to publish. It is which pages already get cited about your category, and what they say about you. Pull the retrieved and cited URLs for your prompts, which is what the Sources view is for. Then work down the list.

• Roundups that omit you. Most publish an inclusion process. Use it. A paragraph on a page ranking for six variants beats a new page of your own ranking for none.
• Listings that describe you wrongly. Old pricing, a limitation you removed, a feature you shipped last year. Answers repeat these verbatim. Fixing it is usually one email.
• Review directories. A thin profile gives a summariser nothing to quote.
• Threads about your category. Take part as yourself, visibly.

Do not manufacture recommendations. That is not moralising, it is operational. Astroturfed reviews are cheap and detectable. They also poison the corpus you are trying to look good in, and a criticism that lands there gets repeated for years.

These pages are retrievable today and training data tomorrow. It is the one lever that pays into both halves.

Step 4: The tactics that survived a controlled test

Here the advice can stop being folklore, because somebody ran the experiment.

The original GEO paper, from researchers at Princeton, Georgia Tech and the Allen Institute, built a benchmark of 10,000 queries. They rewrote source pages in specific ways, then measured how prominently each version appeared in the generated answer. Their score weights how early and how fully a source gets used. Leaving the page alone scored 19.3.

Rewrite applied to the page	Score
Adding quotations from credible sources	27.2
Adding relevant statistics	25.2
Improving fluency	24.7
Citing sources	24.6
Adding technical terms	22.7
Making the tone more authoritative	21.3
Adding unique words	20.5
Keyword stuffing	17.7

Quotations, statistics and citations gained 30 to 40% over doing nothing. The authors repeated the test on Perplexity and saw gains up to 37%. Keyword stuffing, imported from SEO, scored below leaving the page alone.

Now look near the bottom. Writing in a more authoritative tone barely moved. The authors conclude that generative engines “are already somewhat robust to such changes”. The most repeated advice in this field is the one the controlled test found closest to useless.

Two limits, stated plainly. The paper’s engine design, in the authors’ words, “closely resembles the design of BingChat” at the time of writing, so it predates today’s engines. And it measures prominence in an answer, not sales. It is still the best controlled evidence in public.

What follows for your pages:

• Put checkable specifics on them. Figures, dates, prices, named limits, quoted sources. Given three pages saying the same thing, the one with a real number survives the summary.
• Answer in the first sentence of the section. One claim per paragraph. What gets lifted is a passage, not a page.
• Name the qualifier. Fan-out queries carry them: team size, country, budget, industry. A page that names one beats a general page.
• Write the comparison pages. Buyers ask assistants to compare constantly. Without an honest answer on your site, the engine uses someone else’s.

On freshness, the quoted claim is that over 70% of cited pages were updated in the last year. Alone that proves nothing, because people mostly ask about current things. The Seer Interactive study separates the two dates. By last update, 72% of cited content looks current. By original publication, only 42% is under a year old. More than a quarter of fresh-looking citations are old pages that were maintained. The instruction is maintenance, not churn.

Step 5: Let the fetchers in

Boring, and the only place where one config line costs you the whole channel. Every vendor publishes which agent does what. The distinctions matter more than most robots.txt files reflect.

Agent	What it does	Blocking it costs you
GPTBot	OpenAI’s training crawl	A place in future models
OAI-SearchBot	The index behind ChatGPT search	Appearing in ChatGPT search answers
ChatGPT-User	Fetch triggered by a user’s question	The live read
PerplexityBot	Perplexity’s index	Perplexity citations
Perplexity-User	Perplexity’s user-triggered fetch	The live read
ClaudeBot	Anthropic’s training crawl	A place in future models
Claude-SearchBot	Anthropic’s search indexing	Visibility in Claude’s search results
Claude-User	Claude’s user-triggered fetch	The live read
Google-Extended	A robots.txt switch, not a crawler. Covers Gemini training and grounding	Gemini only. Google says it “does not impact a site’s inclusion in Google Search”, so it does not remove you from AI Overviews

Three things follow, and none is obvious.

Training and search are separate choices. OpenAI supports allowing OAI-SearchBot while blocking GPTBot. Appear in search, stay out of the training set. Blanket disallow rules written during the 2023 scraping panic block both, and are still live on a lot of sites. Allow about 24 hours for a change to reach OpenAI, and the same for Perplexity.

The fetch that matters most may ignore robots.txt. Because a person asked for it, OpenAI says robots.txt “may not apply” to ChatGPT-User. Perplexity says Perplexity-User “generally ignores robots.txt rules”. So robots.txt is not where an accidental block happens at answer time. Your firewall is.

Cloudflare changed the default. Since 1 July 2025, every new domain is asked whether to allow AI crawlers, and starts from blocked. If your site was set up after that date, check rather than assume.

Then run curl against your key pages. If the content only appears once JavaScript has run, a live fetcher sees an empty shell.

Step 6: The slow half

If the diagnostic pointed at memory, the levers change and so does the clock. Nothing here pays off this quarter.

What moves it is being written about, widely, on sources that get absorbed. Press, analyst coverage, editorial placements, reference works, documentation other people quote. Feedback arrives in training cycles, not weeks. The work you commission now aims at a snapshot nobody has taken yet.

The compensation is that this half is cheap and steady to measure, because nothing browses. With tools off, ask a model what it knows about you, who your competitors are, what the common criticisms are, and how it would score a list of brands on authority. Name the geography every time. Repeat quarterly. The full version is in the half of AI visibility that is not SEO.

What not to spend money on

• llms.txt. Google says it does not use it, and John Mueller compared it to the keywords meta tag. That is about as damning as that team gets. Publish one if you like. Nobody should bill you for it.
• Schema as the lever. Structured data earns its place where a result type needs machine-readable fields. On an ordinary article it restates the title and the date. We went through this at length.
• Keyword stuffing. Measured at 17.7 against a 19.3 baseline. You are paying to do worse than nothing.
• Tone-of-voice rewrites sold as GEO. Measured at 21.3, the weakest thing that worked at all.
• Hidden text and prompt injection. It does not survive, and it is how a domain gets dropped from a retrieval pipeline for good.
• One-off audits. One snapshot of a system with a 2.8 point spread is a screenshot, not a measurement.

The objection worth taking seriously

On a Hacker News thread about an AI visibility product, ceejayoz put the case against this whole category in one line: “Ugh. The worst of SEO, but a bunch more of it? Noooooo.”

The founder argued that engines will reward genuinely useful content, because that is what makes them useful. The reply sharpened:

the appearance of useful to Google is more important than the actual being useful to Google’s visitors.

That is the right worry. No vendor here, us included, gets to wave it away. The honest answer is that it lands hard on tactics that optimise appearance, and much less on the ones that survived the test. Adding real statistics, real quotes and real citations makes a page more useful and more quotable at once. Rewriting it to sound authoritative only makes it sound authoritative, and it scored near the bottom. That gap is the whole argument. Choose which side your budget lands on.

Knowing whether it worked

Two habits separate this from guessing.

Do not read a move smaller than your noise. Learn what normal variation looks like before you call a change a result.

Hold something back. Apply a tactic to every page you care about and you will never know if it did anything. Comparable pages, the change on some, withheld from the rest, both tracked for months. It is uncomfortable to propose. It is also the only thing that settles an argument here.

Be clear which column you are moving, too. Retrieved, cited, mentioned and recommended are four different outcomes. Two tools reporting different numbers are usually counting different columns.

The short version

1. Diagnose whether you are unknown or merely unfindable.
2. Find the fan-out queries. You are not competing for the prompt.
3. Fix the pages you do not own before writing new ones you do.
4. Add statistics, quotes and citations. Skip the tone rewrite.
5. Check the fetchers are allowed in, at the firewall as well as robots.txt.
6. Start the slow work now. Its lead time is measured in training cycles.
7. Re-measure daily, and hold a control group back.

Steps 2, 3 and 7 need the same thing: knowing what the engines retrieved and cited for your prompts every day, not once. That is what Voxoria does. Daily runs across ChatGPT, Perplexity, Gemini and Google AI Overviews, the fan-out queries behind each prompt, who got named, which sources were retrieved and cited.

We sell this, which is a reason to check the sources rather than take our word for it. Every claim above links to the documentation or the paper it came from.