Your domain did not appear among the sources for generic queries

This is the only check in the whole audit that judges not your site but the outcome. The module puts a series of generic questions to language models with web access (questions without your brand name in them, the kind a customer who does not yet know you would ask) and looks at whether your domain appeared among the sources the model answered from.

The questions go through the maker's API, not through its app. That difference matters: this is not the same as what you see in your own ChatGPT or Claude. The app adds its own instructions for the model, rewrites your question its own way and takes account of memory, location and your sign-in, so two people asking the same thing get different answers. Through the API the question arrives without any of that history, which is the only way to get a repeatable and comparable number.

Who does the searching differs by model. For some models the maker's own search tool finds the sources. For others the searching is done by the search engine of OpenRouter, the service we call the models through, and the model is handed its results. In that second case we are not measuring what the model would find on its own. The report says which of the two applied for each AI.

The finding is raised only when both conditions hold: there was at least one valid generic question (no error) and in none of them did the domain appear among the sources. When there are no valid generic questions at all, the check stays quiet; “never cited” would be an unsupported claim without them.

A domain counts as cited when it is among the sources the model used, not only when it is named in the answer's text. That is the gentler and more honest measure.

What the check does not do:

• It does not ask every model. It is a few specific models with search, not a survey of the market.
• It does not measure the answer in your app. It is a clean API question: no app instructions, no memory, no location, no sign-in.
• It does not repeat questions for a second opinion. Each is asked once. The one exception: if the model searched but returned no source at all, the question is asked once more, and if sources still do not come back it is recorded as not measured and left out of the score.
• It does not cover brand questions in this finding; the module tracks those separately and they do not trigger it.
• It will not tell you why you are not cited. That is the rest of the report's job.

The finding is site-level and informational; it costs no score at all. The next section explains why.

This finding needs a different explanation from the others, because it is not a recommendation with evidence; it is a measurement of an outcome. And that measurement has limits you must know before drawing a conclusion from it.

Limit one, and the most important: a language model is not deterministic. Ask the same question twice and you may get two different sets of sources. Our result therefore does not say “your domain will never appear”, it says “with these questions, at this moment, with this model, it did not”.

Limit two: the sample of questions is small, and phrasing changes the outcome substantially. A different wording of the same question can return different sources.

Limit three: not appearing for a generic question is normal. For a broad question a model typically reaches for three to five sources from the entire web. Not being among them is no mark of a defect; it is the default state for the overwhelming majority of sites in the world.

Limit four, and easy to miss: we measure the API, not the app. If you ask the same question in your own ChatGPT and get a different answer than we did, neither side is wrong. The app wraps the model in its own instructions, its own rewrite of your question, memory and your location; we ask without any of that, because otherwise the number could not even be compared with itself a month later.

That is exactly why the finding is informational and costs no score. Deducting points for not winning a lottery with five places would be dishonest.

So what should you do about it? The finding itself will not name a cause. But the direction to look in is documented, and it is the one place in the field with a peer-reviewed experiment: Princeton (KDD 2024, 10,000 queries) measured that visibility in generated answers was lifted by quotations in the text (+41 %), concrete statistics (+33 %) and links to sources (+28 %). The last of those is measured by AUTH-01; the first two we do not measure at all.

Expectations first: this finding is not fixed by a single change and it may take months for the outcome to shift. Anyone promising otherwise is selling something that cannot be documented.

What is worth doing, ordered by strength of evidence:

1. Confirm anybody can read you at all. ACC-01, ACC-03 and ACC-07 are necessary conditions; while they are open, the rest is moot. A site a crawler cannot enter will never be cited. Treat ACC-05 (content behind a login) the same way. It is a warning because it rests on a heuristic, but a crawler cites nothing it cannot read.
2. Put concrete numbers in the text instead of general claims. Nobody cites “fast delivery”. “Delivered within two working days on 94 % of orders” they do, because it is an answer, not a promise.
3. Add quotations and links to sources (AUTH-01). A claim with a traceable origin is more usable than one without.
4. Write content for the questions people actually ask. Look at the generic questions the audit used, and ask whether your site answers them anywhere at all.

And one thing you would not hear from us elsewhere: it is possible that everything is in order and you still will not be cited. For generic questions you compete with the whole internet for five places. For most sites the more sensible goal is appearing for specific, narrow questions in your field rather than the broadest ones, and there the odds are incomparably better.