Ranking and being cited are not the same thing

If my page ranks well, why is the AI answer citing someone else?

Because ranking makes you eligible and retrieval decides. Google's stated bar for appearing as a supporting link in AI Overviews or AI Mode is that a page is indexed and eligible to be shown with a snippet - the ordinary bar. It then adds the sentence that every site owner should internalise: meeting every requirement "doesn't mean that Google will crawl, index, or serve its content."[^google-eligibility] Eligibility is a filter. Selection is a separate step, and it runs against a question you did not see.

The query you optimised for may never be run

Google says its AI features may use a "query fan-out" technique - issuing multiple related searches across subtopics and data sources to develop one response.[^google-fan-out] A user asks "what's the best AI visibility tool for a small SaaS team". The system decomposes that into sub-questions: what tools exist, what do they cost, which support small teams, what do users complain about. Each sub-question runs its own retrieval. The answer is assembled from the union.

Three consequences follow, and they are why keyword-shaped thinking breaks here.

Your competitor set changes per sub-question. You may be the best page in the world for "AI visibility tool" and have nothing that answers "what do users complain about", which is the sub-query that decides how the answer describes you.

Breadth beats a single perfect page. Google says that while a response is being generated, its models identify more supporting pages, "allowing us to display a wider and more diverse set of helpful links associated with the response than with a classic web search."[^google-more-diverse-links] More slots, distributed across sub-topics. A site with eight adequate pages covering eight facets can out-cite a site with one excellent page covering one.

The head term is not the unit of work any more. The unit is the question, and questions come in families. This is why the useful artefact for an AI visibility programme is a prompt set - the twenty questions a buyer actually asks - rather than a keyword list ranked by volume.

Rank helps, until it does not

Here is the most counter-intuitive published result in this field, and it is worth quoting carefully because it is easy to over-read.

The GEO study measured its content modifications separately for sources at each starting rank in the underlying search engine. Adding citations to the page improved visibility by 115.1% for fifth-ranked sources and reduced it by 30.3% for first-ranked sources.[^geo-rank-asymmetry] The authors' conclusion: the technique is "especially helpful for lower ranked websites."

What that does not mean: that ranking first hurts you. The first-ranked source already dominates the answer, and the paper's metric is a share of a fixed pie - position-adjusted word count, which weights each source's contribution by an exponentially decaying function of where its citation sits in the answer.[^geo-visibility-metric] Restructuring a page that is already the primary source can only redistribute attention away from it.

What it does mean, practically: if you are not the incumbent, structural content work is the highest-leverage thing available to you, and the incumbent cannot copy your way out of it, because the same edits do less for them. This is the rare case where the challenger has the better return on effort.

Mention, citation, and the gap between them

Three different things get called "AI visibility", and a programme that does not separate them will report progress it did not make.

You can be mentioned without being cited, which usually means the model learned about you from somewhere it is not linking. You can be cited without being recommended, which usually means your page answered a factual sub-question while someone else's answered the buying one. And you can be mentioned inaccurately with a competitor's comparison page as the source, which is the worst outcome available and completely invisible to anyone counting mentions only.

Track all three separately, per question. A single "visibility score" that blends them tells you nothing you can act on.

Why the same question gives different answers

Two structural reasons, both documented, before you reach for model non-determinism as an explanation.

Surfaces differ. Google says AI Mode and AI Overviews may use different models and techniques, "so the set of responses and links they show will vary."[^google-different-models] Two Google surfaces, two answers. The same is true of a consumer assistant versus the same vendor's API.

The feature may not fire at all. Google says AI Overviews are shown only when its systems determine they are additive to classic Search, "and as such, often don't trigger."[^google-ai-overviews-trigger] An empty result in your tracking is frequently not a lost citation; it is a query where no AI answer was generated. Recording those two states as the same thing manufactures a decline.

This is why measurement needs a sampling design rather than a screenshot, which is chapter six.

What replaces rank as the number to watch

Rank is not useless. It is a cheap, dense signal that your page is retrievable, and Google's AI features draw on the same index, so a page that cannot rank generally cannot be retrieved either. Keep it as a diagnostic. Retire it as the objective.

The replacement is a small set of questions and a rate:

1. Define the prompt set. Ten to thirty questions a real buyer asks, in their

words, including the unflattering ones ("X vs Y", "is X worth it", "alternatives to X").

1. Sample each question repeatedly, across the platforms you care about, and

record for each run: were we mentioned, were we cited, were we recommended, and which URLs were cited instead.

1. Report rates with sample sizes. "Cited in 4 of 14 runs" is a measurement.

A bare percentage with no sample size behind it is a decoration.

1. Watch the competitor URL list, not just your own rate. The pages cited

instead of yours are the specification for what to write next; they are the answer the model preferred.

That fourth item is the one most teams skip and the one that pays. If the same review site is cited on eight of your twenty questions, your next move is not a blog post - it is getting your entry on that review site accurate and complete, which is chapter four.

Diagnosing which step you are failing

"We are not showing up in AI answers" is four different problems wearing one sentence. The symptoms separate cleanly, and the diagnosis takes an hour.

Most teams assume retrieval and are actually failing access or corroboration - the two ends of the chain, and the two nobody owns. The middle two get all the blog posts because they are the two a content team can act on alone.

One caution on the retrieval test: an assistant asked a leading question will often produce a flattering answer that no real buyer would ever see. Ask the question the way a buyer asks it, including the version where your brand is not in the prompt at all. The unprompted version is the one that matters, and it is usually the one that hurts.

Why the winner of the classic result is often not the source

One more asymmetry worth naming, because it changes who should bother with this work at all.

Classic search rewards the page that best matches the query. An answer rewards the passage that best completes the sentence the model is about to write. Those are different objects. A category-leading homepage matches "AI visibility tool" perfectly and contains almost nothing quotable about how such tools differ. A mid-sized competitor's comparison page contains three quotable sentences about exactly that, and gets cited in an answer whose first line names the leader.

That is the everyday shape of AI answers in software categories: the leader is mentioned, someone else is cited. Mention accrues to whoever the model already associates with the category; citation accrues to whoever wrote the most useful passage this week. The first is slow to change and the second is not, which is the entire reason a challenger has anything to work with here.

It also explains an outcome that confuses people the first time they measure it - appearing in answers constantly while receiving no citations at all. That is not a retrieval failure. It is being famous and unquotable, and the fix is in the next chapter.

What to take from this chapter

Being fetched gets you into the candidate pool. Ranking gets you into the candidate pool faster. Neither one writes you into the answer. The gap between "eligible" and "cited" is filled by whether a passage on your page is the most useful available text for one of the sub-questions the system generated - which is a content-structure problem, and the subject of the next chapter.