Skip to content

What AI Assistants Cite

On technical questions, assistants reach for primary standards documentation rather than agency commentary, and often cite nothing at all. Across 18 neutral technical queries run with personalization off, W3C was cited on 3, Google for Developers on 4, and 11 answers cited no source whatsoever.

What we measured

Baseline wave, recorded 2026-09-17 against a query set frozen and hashed the day before. Outcomes on the five-level ladder: 0 companion page citations, 0 domain citations, 0 video citations, 0 unlinked mentions, 18 no-appearance. Of the sources the assistant did cite, W3C appeared on 3 answers, Google for Developers on 4, and 11 answers cited nothing at all. A first attempt on a personalized account was discarded: it returned our own brand and a client as example content, which no neutral session produces. That run is kept on disk, marked contaminated, rather than deleted.

Why it matters

  • It identifies the field you are actually competing in. On a definitional question about a web standard, the competitor is the standards body, and that is not a competition worth entering.
  • The 11 answers that cited nothing are the opening. An answer with no source attached is one where no source was good enough to attach.
  • It sets a realistic ceiling. Nobody is going to displace W3C on what an unlabeled link is, and a content plan that assumes otherwise will fail quietly for months.
  • It explains what to lead with. Original measurement is the only material here that a standards body does not already have, because nobody else ran it.

The failure

The failure: a page that restates what the standard already says
# What Is a Canonical URL?

A canonical URL tells search engines which
version of a page is the preferred one. You
set it with a link element in the head:

    <link rel="canonical" href="...">

This helps avoid duplicate content issues
and consolidates ranking signals.

Accurate, well written, and with nothing in it that Google's own documentation does not say better and with more authority. On our baseline, questions of exactly this shape were answered with a citation to Google for Developers. There is no reason for an assistant to reach past the source for a restatement of it.

The fix

The fix: lead with the measurement nobody else has
# Canonical URLs

We audited canonical tags across 217 pages of
our own site and found 9 self-referencing tags
on noindex pages, which every tool we tried
reported as a defect and none of them are.

Here is the audit script, the output, and why
the tools disagree with the specification.

Same topic, different claim. The first version competes with the specification and loses. The second carries a number, a denominator and a reproducible script, none of which exist anywhere else. That is the only durable advantage available on a definitional query.

Verify the fix

Changing the code is not the same as fixing the problem. Confirm it.

  1. Freeze your query set before you publish anything, and hash the file. A query reworded later is a different question, and comparing across a reworded set measures nothing.
  2. Run logged out, or with memory and custom instructions off. Our first attempt returned our own brand on generic questions, which is account personalization rather than retrieval.
  3. Check the personalization check every wave. Generic questions should return generic examples. If your own names appear, discard the run.
  4. Record what was cited, not only whether you were. What the assistant reaches for instead is more actionable than your own absence.
  5. Compare at constant page age rather than on one calendar date. A page published in March and a page published in August are not comparable on the same afternoon.
  6. Record the model version. Treating two model versions as the same system is how a platform change gets read as a content result.

Exceptions and misconceptions

Zero citations at baseline is the expected result

The pages were an hour old. A baseline exists to be compared against, not to be interpreted. Reading a pre-publication zero as evidence about page quality is a misuse of it.

One assistant is not all assistants

This baseline is ChatGPT with search only. Retrieval differs by assistant and by model version, and a result here does not transfer to another product without measuring it there.

A URL in a sources panel is not a citation

It counts only when the assistant uses it to support the response. Counting panel appearances inflates results and makes waves incomparable once a product changes how the panel works.

Nobody can promise inclusion or citation

Not us, not an SEO vendor, and not the companies operating the assistants. OpenAI's own publisher guidance stops short of it. Anyone selling guaranteed AI citations is selling something that does not exist.

Watch it

Primary sources

The measurement above comes from our study, Learn Hub AI Citation Protocol, baseline data.

Related lessons

Check your own site for this

DeepAudit AI renders your page in a real browser and reports the affected code, so you can see exactly where each finding came from. Free, no signup.

Run a free audit

Last reviewed 2026-09-18. Checks covered: .