ChatGPT named you in 0 of 12 buyer questions. The answer is in the 300 sources it did use.

Why is your brand not showing up in ChatGPT answers? Sort the 300 sources it read into four failure modes, run one cheap test, and fix retrieval first.

Last updated:

15 min read

Summary

  • Retrieval explains 32 percent of citation variance within a site while all on-page properties combined explain just 2.8 percent — which is why a backlink, on its own, often changes nothing.
  • Backlinks are not wrong, just wrong in sequence: if the engine never retrieves your page, there is nothing for a link to act on, so fix retrieval, then the answer, then authority.
  • The source list is the diagnostic — sort it into four buckets (competitor pricing, "best X" roundups, Reddit threads, only your own pages) to find which failure mode you are in: not crawled, not retrieved, retrieved but not named, or named without citation.
  • Run the cheapest possible test: one page titled after your highest-intent question, with price, best-for, and problem solved stated in the first 1,000 characters, then re-check your prompt set over several days — 75 percent of AI answers change daily, so measure frequency, not a single snapshot.
  • Mole's AI SEO & GEO Agent runs that loop for you — prompt-set visibility tracking, Reddit listening for real buyer language, research-backed content, and indexing — so you are measuring the source mix instead of guessing at it.

You ran twelve buyer-intent questions through ChatGPT with web search on. Questions like "best X for small teams" and "X vs Y for enterprise." Your company appeared in none of the twelve answers. Then you checked the source list — roughly 300 pages the engine actually read — and your domain was not in it at all.

Somebody in a forum told you to build more backlinks. That advice is not stupid, but it is almost certainly not first. Links are slow and expensive, and they may be fixing a problem you do not have.

ChatGPT handed you a diagnostic report. Every page in that source list is one its internal search retrieved and read before it wrote a word. The list shows where the engine went instead of going to you, which is the same thing as showing you what to fix. The method is in the list.

Sort the source list into four buckets

Open the source list and sort it before you touch anything else. Every source type maps to a different failure mode, and each one takes a different fix.

Source type in the list What it means What to do
A competitor's pricing, plans or feature page The engine ran a sub-query about that competitor and never looked for you Put your own facts in plain text: the actual price, the plan contents, and the features named the way buyers name them. A "Contact us for pricing" page cannot be retrieved.
A "best X" roundup listing three or more competitors The shortlist was decided on someone else's page before your site was read Get into independent roundups. Also publish your own honest "alternatives to X" pages — vendor-written comparisons do get cited for comparison questions.
Reddit threads The engine is sourcing community opinion and real-world language, not your spec sheet Be present in those exact threads with genuinely useful answers. The language in them is also free copy research.
Only your own pages You are retrievable — the engine found you. The problem is the answer or its order on the page Put the key comparison facts (who it's for, what it costs, which use case it fits) in the first block of text. State them plainly.

A source list full of competitor pricing pages means the engine chose its shortlist before it searched for anyone else. One vendor's reading of a leaked ChatGPT pipeline — which nobody outside that vendor can replicate — suggests that roughly 24 to 30 percent of the engine's internal fan-out queries take the form of a brand name plus the word "official." The engine decided who to investigate before it started comparing. If your brand was not on that starting list, it was never in the race.

A Seer Interactive analysis of ChatGPT's reasoning steps found that fewer than one in ten brands get named before the engine runs a single search. Most, 59 percent, arrive only after that first round of broad searching. Those searches are where roundups decide the shortlist: if your brand is not on the pages that round lands on, the list is closed by the time the engine looks for specifics.

Reddit threads in the source list mean something else again. The engine is sourcing opinion and real-world language, not your spec sheet, and the same analysis notes that, in ChatGPT, Reddit is the source it leans on for recommendation and community-opinion queries. Do not fight that — join the forum. The threads also tell you how buyers phrase the problems your product solves, in words that are usually uglier and more specific than anything in the category's marketing. That language belongs on your pages.

If only your own pages appear in the source list, you are in the best position available. Retrieval is solved. The problem is what the engine found when it read them.

Domain authority does matter. It just matters downstream of retrieval, and if the engine never retrieves your page there is nothing for a link to act on.

The largest controlled study of this published to date analyzed 158,099 answers and 468,037 citations to separate what actually drives citation rates (Chen and Koudas, 10 October 2026; associations measured through the APIs rather than consumer apps, two industries, English, US questions). The findings are direct:

  • Retrieval explains 32 percent of citation variance within a site.
  • All on-page properties combined explain 2.8 percent.

Within a site, only two page-level features survived the controlled test: a title that matches the question, and an opening paragraph that matches the question. FAQ blocks, structured data, author boxes, summary boxes and page speed all washed out. The study's structural point is that each engine's own search rarely returns what Google or Bing would for the same prompt, so the levers you learned for one surface do not transfer cleanly to the other.

That does not make authority irrelevant. In the same vendor's replay experiments, identical page text was cited in 20 of 20 answers on an established publisher's domain and 0 of 20 on a Blogspot address. But a page carrying the exact fact the question needed was cited almost every time regardless of domain, and an off-topic page almost never was, even on the strongest domain. Authority decides close calls. It does not get your page into the retrieval pool, and it does not beat a page that answers the question.

What makes this worth checking is how consistently the sequence gets skipped. When one founder posted exactly this result — twelve questions, twelve misses, no page of theirs among roughly three hundred sources — two of the first replies were "Those [competitors] have more authority in search results. Get more backlinks" and "Authority issue." The one comment that started from the source list, splitting it by type and asking what sub-question each competitor page had answered, sat far below them.

That is not a knock on the people answering. "You need authority" is the right answer to a lot of SEO questions, so it is the reflex. But it is a claim about your domain, and the source list is a claim about your pages. Only one of those is cheap to test.

Still Not Retrieved? Mole's agent diagnoses your retrieval gaps and fixes them — fast. Book a demo.

Four failure modes and the cheapest test for each

Once the source list is sorted, the diagnosis usually lands in one of four states. Work through them in order.

  1. Not crawled. Check your server logs for the OpenAI search crawler. Blocking the training crawler and blocking the search crawler are two different decisions, and only the second takes your pages out of reach of being cited at all. Confirm robots.txt allows the search one, and that the site is reachable.

  2. Not retrieved. Your domain does not appear in the source list for any of the twelve questions. The engine built its shortlist from competitor pages and third-party roundups, and your name was not among the candidates it searched. The fix: get the facts onto your own pages, and get your brand onto the independent lists the engine already reads.

  3. Retrieved but not named. Your domain appears in the source list, but your brand does not appear in the final answer. The page was read and the answer was buried. The leaked-pipeline analysis found that getting a brand's positioning wrong — describing a budget tool as enterprise-grade, for example — dropped citation rates by 44 to 95 percent in controlled tests, and that a wrong price for a price-sensitive brand dropped them by 100 percent. The fix is ordering and accuracy: put the right facts first, and state them plainly.

  4. Named without citation, or cited without being named. These are two separate events, and both directions happen. The engine sometimes draws on parametric memory (what it learned during training) and names a brand without citing a source. It also sometimes cites your page as evidence for a claim about a competitor. Track them separately. Counting brand mentions and counting source appearances are not the same measurement.

The smallest test that confirms a diagnosis

You do not need a full content calendar to confirm whether this works. You need one page and one question.

Pick the highest-intent buyer question you are currently invisible for. Write a page whose title matches it closely. In the first 1,000 characters of the page body, state the facts a comparison needs: who this product is best for, what it costs, and what specific problem it solves. Do not bury these facts below a hero section or a features grid. State them first, in plain text.

Get the page indexed. Then re-run your original twelve questions — repeatedly, over the next few days and weeks. Indexing is fast; whether the answer moves is not a one-off event, and how long it takes varies by category.

Two rules for reading the results honestly.

Measure frequency, not a single snapshot. AI answers shift day to day. Across a sample of tracked prompts, 75 percent of answers changed from one day to the next, and only 42 percent of the pages cited on day one were cited again on day two. A single before-and-after screenshot proves nothing. Run the same question set multiple times across several days and compare the frequency of appearances, not the result of one run.

Judge the source mix, not just whether you were named. The thing you control is whether your page appears in the list of sources the engine retrieves. That is the leading indicator. Brand mentions in the final answer are a downstream outcome. If the source mix changed — if your domain now appears where it was previously absent — the fix is working, even if the name has not surfaced consistently yet.

The comparison question is the fastest entry point. If your source list contains competitor pricing or feature pages, buyers are already asking "X vs Y" and the engine is reading your competitor's answer to it. Put your own answer on a page whose title matches the question. State the price. State what each plan includes. Name the features the way buyers name them, not the way your internal team does. The Seer analysis found that while only 13 percent of ChatGPT's first-round searches target a specific website, 75 percent of later-round searches do — using the site: operator on domains already identified as relevant. Get identified as relevant in round one, and round two searches your site directly.

Running the loop

The process above is not a one-time fix. It is a loop: pull the buyer questions from real language, run them through the engine, sort the source list to find the failure mode, write the page that answers the sub-question the engine was actually running, get it indexed, re-run, compare the source mix.

Mole is an AI SEO & GEO agent built to run that loop. The questions come from what buyers actually wrote — in Reddit threads the engine already cites, and in your own sales calls — not from a prompt list someone guessed. The page that answers the sub-question gets drafted against one set of approved product facts, so the price and the use case read the same everywhere, and it goes through an editorial gate before it ships. Then it is indexed, and the same questions run again on a schedule. The diagnosis, the writing and the measurement are one workflow, not three tools.

Run the Loop for You — Mole executes your GEO strategy — tracking, content, and indexing in one agent. Get the agent.

Go back to the forum thread where someone told you to buy links. The founder there had the answer the moment they looked at the source list — three hundred pages, not one from their domain. That is not a backlink problem. It is a retrieval problem, and the list said exactly where the engine went instead.

The advice to buy links was not wrong in principle. It was wrong in sequence. You cannot buy your way into a source list the engine never opens.

Look at the list before you spend. The engine is showing you its work.

Frequently Asked Questions

What is a source list in a ChatGPT answer?

A source list is the set of pages ChatGPT's internal search retrieved and read before composing an answer. When you run a prompt with web search on, the tool surfaces the roughly 300 pages the engine actually consulted. That list is a diagnostic: every entry shows where the engine went instead of going to you, and that is where you start.

Yes, but they matter downstream of retrieval, not first. If the engine never retrieves your page, there is nothing for a link to act on. A large controlled study found retrieval explains 32 percent of citation variance within a site while all on-page properties combined explain just 2.8 percent. Fix retrieval first, then the answer, then authority.

Why is ChatGPT not mentioning my brand?

Usually because the engine never retrieved your page, or it read the page but your key facts were buried. There are four distinct failure modes: not crawled, not retrieved, retrieved but not named, and named without citation. Each has a different fix, so you have to diagnose which one you are in before you write or build anything.

How do I find out why AI search engines ignore my site?

Sort the source list into four buckets: competitor pricing pages, "best X" roundups, Reddit threads, or only your own pages. Each bucket maps to a different problem. Competitor pricing pages mean the engine built its shortlist before searching for you. Reddit threads mean it is sourcing community opinion. Only your own pages is the best sign — retrieval is solved and the issue is on-page order and accuracy.

What should I put in the first 1,000 characters of a new page?

State the facts a comparison needs: who the product is best for, what it costs, and the specific problem it solves. Put them first, in plain text, before any hero section or features grid. A title that matches the question and an opening paragraph that matches the question are the only page-level features that survived a controlled test within a site.

How long does it take for AI visibility to improve?

Re-run your prompt set repeatedly once the page is indexed — over several days and then weeks — rather than picking one date. How long a shift takes varies by category. And measure frequency across those runs rather than a single snapshot: about 75 percent of AI answers change from one day to the next, so one before-and-after screenshot proves nothing.

How often do AI answers change?

Roughly 75 percent of AI answers change from one day to the next, and around 42 percent of pages cited on day one are gone from the citation list on day two. That is why you track the frequency of appearances across a question set, not the result of any single run.

Should I track brand mentions and citations separately?

Yes, they are two separate events and both directions happen. The engine sometimes names a brand from its training data without citing a source, and sometimes cites your page as evidence for a claim about a competitor. Counting only mentions or only source appearances gives you an incomplete picture, so measure both.

How do I choose a tool to track and improve AI visibility?

Look for one that runs the whole loop — prompt-set visibility tracking, real buyer-language research, content production, and indexing — rather than a single-slice tracker. Mole's AI SEO & GEO Agent is built for exactly that: its AI Visibility Tracker shows where answers mention, cite, or omit you, its Reddit listening surfaces real buyer phrasing, its content writer turns the diagnosed gap into a sourced page, and its Google Indexer submits that page the moment it goes live, at up to 200 URLs a day. If you would rather run the loop without building it yourself, get in touch.