Why AI answers describe your category but never link to you

“Why does ChatGPT not link to my website?” Usually because the answer was built from other people’s pages, not yours. In LeadsNow’s log of 899 answers from ChatGPT’s API with web search (not the consumer app) to 105 buyer prompts, 3 July–23 September 2026, it linked us on just 12 prompts and named us without a link only 3 times.

• Three gates, in order: did the engine search at all, was your page among the sources it consulted, and did it link your page rather than a roundup that describes you.
• Most common failure: the answer links sources, just never yours (gate 2 or 3). That was 93 of our 105 ChatGPT prompts. AirOps’ vendor analysis of 21,311 brand mentions put third-party domains at 85%.
• Rare failure: being named with no link. 3 of the 112 ChatGPT answers that surfaced us at all.
• Link-free answers: 30 of 899 ChatGPT answers (3.3%) linked nobody at all.
• Cheapest check: make sure robots.txt does not block OAI-SearchBot, and that Google can show your page with a snippet.

Why does ChatGPT not link to my website when it describes my category?

An AI answer that describes your category without linking you has failed at one of three gates. We call this the Search–Consult–Link test. Each gate has a different cause and a different fix, so the first job is to work out which gate you are failing at.

1. Search. Did the engine retrieve anything? OpenAI’s web search documentation says that with tool_choice: "auto", “search is optional”. The model can answer from its training data and cite no sources at all.
2. Consult. If it did search, was your page among the pages it read? According to the same documentation, the sources field lists every URL the model consulted, and “the number of sources is often greater than the number of citations”.
3. Link. If it read your page, did it credit your page? Or did it credit a “best X” roundup that says the same thing, with you further down or missing?

The Search–Consult–Link test matters because rewriting your page does nothing for gate 1 or gate 2, and earning mentions does nothing for gate 3.

How often each failure happens: LeadsNow’s own ChatGPT and Gemini log

LeadsNow runs a citation monitor against a registry of real buyer prompts in its own niche, lead generation and appointment setting. Each poll records every source URL the answer cites and whether any of them is leadsnow.ai. The figures below come from one read-only query against that database. Engine and method: “ChatGPT” means OpenAI’s Responses API with the web_search_preview tool and tool_choice set to auto. The model was gpt-5 on 884 of the 899 polls and gpt-4o-mini on 15. That is not the consumer ChatGPT app, and its answers can differ. “Gemini” means gemini-2.5-flash with Google Search grounding. Window: 3 July to 23 September 2026, UTC (last poll 15:09 on 23 September).

In LeadsNow’s ChatGPT log, 109 of the 112 answers that surfaced LeadsNow in any form also linked to it (97%). Being named without a link was the rare case. Being absent from the linked sources was the normal one: on 93 of 105 prompts (89%), ChatGPT never linked us in any poll, although 54 of those 93 prompts were polled only once. This is one domain in one niche, so read it as a direction, not a universal rate. Citations flicker between polls; our page on why a competitor is cited in AI answers and you are not sets out the five-poll rule for deciding when a gap is real.

The diagnosis table: symptom, gate, test, fix

The rows are ordered by how often each failure appeared in the log above. The test column is the step that separates one cause from the next.

Did the AI search at all, or answer from memory?

An answer that links nobody was not built from pages, so no change to your page can get you into it. OpenAI’s documentation says the model decides whether to search when tool_choice is auto. Search is forced only with "required". In the LeadsNow log, 30 of 899 ChatGPT API answers (3.3%, 3 July–23 September 2026) came back complete with no source URL at all. There is a limit on that figure: the monitor stores only the first 8,000 characters of each raw response, so it cannot prove that search was skipped in every one of those 30. What it does show is that no page was credited.

A gate-1 miss is diagnosed, not fixed. Log the prompt as link-free and take it out of your citation-rate denominator so it does not read as a page failure.

Was my page among the sources the AI read?

Our own log cannot separate gate 2 from gate 3, because it records citations, not the consulted list. Third-party data points to gate 2 as the bigger of the two for category questions. AirOps’ vendor analysis of 21,311 brand mentions across 500+ commercial-intent queries on ChatGPT, Claude and Perplexity found that 85% of brand mentions came from external domains and 13.2% from the brand’s own site. That is AirOps’ own finding on its own sample, not an independent study.

Before you work on third-party pages, rule out the mechanical exclusions:

• OAI-SearchBot versus GPTBot. OpenAI’s crawler documentation separates the two. OAI-SearchBot “is used to surface websites in search results in ChatGPT’s search features”. GPTBot crawls content that may be used for training. You can block GPTBot and still be cited. If you block OAI-SearchBot, you will not appear in ChatGPT search answers, although you “can still appear as navigational links”.
• Google’s AI features. Google’s guidance for site owners says that to appear as a supporting link in AI Overviews or AI Mode, a page must be “indexed and eligible to be shown in Google Search with a snippet”. It also says there are “no additional technical requirements”. A nosnippet or max-snippet directive limits what these features can show.

If neither applies, the gate-2 fix is to be present where the engine already reads. That means the roundups, directories, forums and earned coverage in your category. Our guide to earned media as a source of AI citations covers how to do that.

Why does the roundup get the link instead of my page?

At gate 3 the engine has read your page and credited another one. Usually the other page is more extractable: a table, a number, a quotable sentence. The best evidence here is the GEO paper (Aggarwal et al., KDD 2024), led from Princeton. In their GEO-bench of 10,000 queries, a generative engine built on GPT-3.5-turbo answered each query from the full text of the top five Google results. Adding quotations, statistics and cited sources to a page raised its visibility in the answer by up to 40%. Keyword stuffing offered “little to no improvement”.

The setting matters for this diagnosis. In each test, the page being optimised was chosen at random from the five sources already handed to the model. So the GEO paper measures gate 3 only: on-page changes help a page that has already been retrieved, and they cannot put a page into the retrieved set. That is why rewriting a page that fails at gate 2 changes nothing.

For gate 3, give each section a self-contained answer with the subject named inside it, a number the engine can quote, and a table where you would otherwise write a list of adjectives. If you want the full breakdown of absence causes, not only this linking problem, see our page on why a business is not showing up in ChatGPT answers.

What running this diagnosis yourself costs

The diagnosis is free if you do it by hand. The arithmetic: 20 buyer prompts × 5 polls each (the minimum before a miss means anything) × 2 engines = 200 answers to read and log. Assuming about a minute per answer, that is roughly three to four hours per round, or around 40 hours a year run monthly.

The gate-1 and gate-2 tests need the API, where the web_search_call item and the consulted sources list are exposed as fields. That means an API key, a short script and a database. Our write-up on tracking your own AI search visibility documents the prompt registry and logging schema we use.

The parts that break at volume are the rate caps on grounded search APIs and the discipline to poll on a fixed schedule. The tracking method is the easy part. The ongoing gate-2 work, earning placements on third-party pages, is where most of the time goes. That is the work that AI SEO and answer-engine services are paid to do.

Frequently asked questions

Why does ChatGPT mention my company but not link to my website?

Either it drew the fact about you from a third-party page and credited that page, or it answered without searching. In LeadsNow’s own log (899 ChatGPT API answers, 3 July–23 September 2026), this happened 3 times, against 109 answers that did link us. OpenAI’s web search documentation says search is optional unless it is forced.

Does blocking GPTBot stop ChatGPT from citing my site?

No. OpenAI’s crawler documentation says GPTBot crawls content that may be used for training, and OAI-SearchBot surfaces websites in ChatGPT’s search features. Blocking OAI-SearchBot is what keeps you out of search answers. Changes to robots.txt take about 24 hours to apply.

Do I need special optimisation to appear in Google AI Overviews?

Google says no. Its AI features guidance states that there are “no additional technical requirements” beyond being indexed and eligible to be shown with a snippet. Snippet controls such as nosnippet limit what AI Overviews can use.

Will adding statistics to my page get it cited?

Only if the engine already retrieves the page. The GEO paper found visibility gains of up to 40% for pages among the top five sources given to the model in its benchmark. It did not test getting a page into that set in the first place.

How many times should I check before deciding I am not linked?

At least five polls of the same prompt on the same engine, across several days. Citations flicker between polls. In our ChatGPT API log (3 July–23 September 2026), ChatGPT linked LeadsNow on 12 of 105 prompts; on 10 of those 12, some polls linked us and others did not, and the other 2 were polled only once.