Short answer: ChatGPT usually cannot find a brand for one of seven reasons, and size is rarely one of them. Its search agent may be blocked, the important facts may only exist after JavaScript runs or inside PDFs, the copy may contain nothing quotable, nobody else on the web may confirm what you say, or the questions you test may be the wrong ones. SEO still gets your pages crawled and indexed. GEO is the extra work of making them easy to quote and easy to corroborate. Diagnose which of the seven applies before you write anything new.

This is a diagnostic playbook. Each blind spot below comes with a way to check it yourself, and the article ends with a self-audit table you can run in an afternoon.

First, know which ChatGPT you are testing

"ChatGPT can't find us" can mean two different things, and the fixes differ.

  • Answers drawn from search. When ChatGPT searches the web, it relies on an index built by OpenAI's OAI-SearchBot, and on ChatGPT-User, which fetches a page because a person asked about it. If those agents cannot reach you, you cannot appear in those answers, however good your pages are.
  • Answers drawn from what the model already learned. GPTBot crawls content that may be used to train OpenAI's models. Blocking it is a legitimate position on training use and does not remove you from ChatGPT's search results.

The roles are documented in OpenAI's bot documentation. The practical point: OAI-SearchBot is the agent sites most often block by accident while meaning to opt out of training, and that single mistake is enough to make a brand disappear from search-based answers.

Three OpenAI agents, three different consequencesHighlighted: the agent sites most often block by accident.GPTBotWHAT IT DOESCrawls content that maytrain OpenAI modelsWHAT IT FEEDSFuture model trainingIF YOU BLOCK ITNothing changes intoday’s answersOAI-SearchBotWHAT IT DOESBuilds the index behindChatGPT search featuresWHAT IT FEEDSSearch-based answersand their citationsIF YOU BLOCK ITRemoved from ChatGPTsearch resultsChatGPT-UserWHAT IT DOESFetches a page becausea person asked about itWHAT IT FEEDSOne user’s request,in real timeIF YOU BLOCK ITThat user’s requestcomes back empty
Figure 1. The three OpenAI agents and what blocking each one costs. Only OAI-SearchBot and ChatGPT-User affect whether you can appear in search-based answers. Source: OpenAI’s crawler documentation.

SEO and GEO: the same pipeline, judged differently

SEO and GEO want the same outcome, a buyer finding you before a competitor, but they are judged by different systems. A ranking algorithm scores pages. An answer engine picks a few passages and a few sources to repeat.

Traditional SEOGEO
What is evaluatedA page's rank for a queryWhether a passage about you is worth quoting in an answer
Unit of contentLanding pages and articlesSelf-contained passages: facts, answers, comparisons, specs
Trust signalLinks and authorityThe same facts confirmed by sources you do not control
Where you need to existMostly your own siteYour site plus reviews, communities, trade press and directories
What you measurePosition and organic trafficWhether you are named and cited, engine by engine, for a fixed set of questions

One finding from the research literature makes the difference concrete. In the GEO paper accepted to KDD 2024, keyword stuffing offered little to no improvement in generative engine responses, while citing sources, adding quotations and adding statistics performed best, with visibility gains of up to 40%. Our GEO vs SEO guide goes deeper on where the two overlap.

Seven reasons ChatGPT can’t find your brand, and how to check each

Seven blind spots, three questionsCheck them in this order: a fix further right is wasted while one further left is open.Can it read you?1Search agent blocked2Facts after JS or in PDFsCan it quote you?3Nothing quotable6No comparison answersDoes it trust you?4Nobody confirms you5Facts contradict7Absent from cited sourcesHighlighted: the cheapest check and the most common cause of total absence.
Figure 2. The seven blind spots in this article, grouped by the question each one fails. Source: the diagnostic method in this article.

1. The search agent is blocked

A robots.txt group copied from a template can block every AI agent with one rule. More often the block is invisible in robots.txt: a CDN or firewall rule returns 403 to these agents while the file says they are allowed. Test with the real user agent instead of reading the file:

curl -s -o /dev/null -w "%{http_code}\n" -A "OAI-SearchBot" https://example.com/

Anything other than a 200 is worth investigating. Our working robots.txt for AI crawlers shows a file that blocks training while allowing search.

2. The facts only exist after JavaScript, or inside PDFs

Most AI crawlers do not execute JavaScript. If prices, specs or answers appear only after client-side rendering, the crawler gets an empty shell, and a 200 response with an empty body should count as a failure. Facts that live only in PDF datasheets, images or sales calls are just as invisible. Fetch a key page and search for the sentence you care about:

curl -s -A "OAI-SearchBot" https://example.com/product/ | grep -i "warranty"

3. The copy gives a model nothing to quote

Copy built on adjectives gives an answer engine nothing to lift. Assistants quote sentences that answer a constraint: a dimension, a material, a return window, a compatibility. "Dishwasher safe, BPA-free, with a two-year warranty" in a sentence is quotable; "crafted for modern life" is not. Put the direct answer in the first sentence or two under each heading, and name the product in full instead of writing "it". Writing for extraction has before-and-after rewrites.

4. Nobody else confirms what you say

Your site is the authority on your product, but it is not independent evidence. A model is more confident stating a fact it finds in several independent places, and a claim that exists only on your own domain is the weakest form of evidence you can offer, however well it is marked up. Search for your brand together with your category and read what reviews, directories and community threads say about you.

5. Your facts contradict each other across the web

The common failure is not silence but a plausible wrong detail: an outdated price, a market you left, a feature that belongs to a competitor. It usually traces back to one stale external source. Write one plain description of what you are and who it is for, use it everywhere you control, and publish an Organization block with sameAs links to your real profiles. Entity and sameAs explains which links to include.

6. You never answer the comparison questions

Many buying questions are comparative: which option suits a beginner, which is cheapest for a given need, how two named products differ. If you never publish a page that states plainly who you are a good fit for and who you are not, the answer is assembled from other people's pages. A comparison that states its criteria is easier to quote than a list of claims.

7. You are absent from the sources the engine cites

Engines lean on different sources. In Profound's analysis of 680 million citations from August 2024 to June 2025, the most-cited domain in ChatGPT was Wikipedia at 7.8% of all citations, with Reddit second at 1.8%, while in Perplexity Reddit led at 6.6%. Those shares are small, which is the point: citations are spread widely, so look at which domains are cited for your own buyer questions rather than chasing one platform.

The most-cited domains differ by engineShare of each engine’s citations, August 2024 to June 2025. Highlighted: the top domain.ChatGPTWikipedia7.8%Reddit1.8%PerplexityReddit6.6%YouTube2.0%
Figure 3. Each engine leans on different sources, and even the leaders take a small share. Source: Profound, analysis of 680 million citations from August 2024 to June 2025.

A quick self-audit

Run through this table before you invest in new content. Each row maps to one of the blind spots above.

CheckWhat good looks likeHow to verify
Search agent accessOAI-SearchBot and ChatGPT-User receive a 200 with real contentcurl with the user agent; check CDN and firewall rules, not only robots.txt
RenderingPrices, specs and answers present in the raw HTMLFetch with scripts disabled and search for the sentence
Quotable factsSpecs, policies and prices stated in sentences, not only in images, PDFs or tablesRead the first two sentences under each heading out of context
Entity clarityOne description everywhere; exactly one Organization node per pageView source and count Organization nodes
CorroborationReviews, directories and press state the same facts as your siteSearch your brand with your category
ComparisonsAt least one page that says who you suit and who you do notAsk the comparison question yourself and read what is cited
MeasurementA fixed set of questions that do not name your brand, reported per engineRemove any test question that contains your brand name

The last row matters more than it looks. Ask an assistant whether a named company is any good and it will name it. In one of our own runs, 13 of 15 generated questions contained the brand name, and the score read 13% when the honest value was zero. Tracking brand mentions in ChatGPT explains how to set up a question set that does not flatter you.

Then run it as a loop, not a project

Once the blind spots are fixed, the work is a cycle: confirm access, keep facts consistent, publish quotable answers and comparisons, earn corroboration elsewhere, then re-measure with the same questions. AI answers vary between runs even when nothing changed, so treat a movement as real only when it appears in two consecutive runs, and never compare runs made with different engines or question sets. There is no honest fixed timeline for results, and nobody outside the engine can promise a mention. What you control is whether your brand is eligible to be named, and whether you can see the result clearly.

Sources

An earlier version of this article quoted usage, shopping-share and survey statistics from statistics roundup sites, plus a composite case study with an illustrative chart. We could not trace them to a single verified original, so they were removed; every number that remains links to its primary source or comes from our own documented measurement.