Ranking in Google versus being cited by AI

What does an AI answer actually draw on?

Far more pages than a results page shows. Across the 61 questions Attensira tracks for its own brand, the 4,374 answers ChatGPT and Perplexity gave in the 30 days to 4 October 2026 carried 97,149 citations: about 22 links per answer. On one question, "Peec AI alternatives", ChatGPT's 57 answers carried 1,703 citations, close to 30 each.

A classic results page shows ten organic links and a person clicks one or two. An AI answer reads a crowd of pages, takes a sentence from some, a fact from others, and attaches a selection as sources. That difference in mechanics is why ranking and being cited come apart.

Here is the shape of one answer, read verbatim from our workspace. Asked "AI search visibility starting price" on 3 October 2026, ChatGPT replied with a short list of vendors and price points: a free tier from one, a low monthly plan from another, Semrush at its listed price, done-for-you services from a one-time fee. Beneath it sat more than 30 source links. One of them was our pricing page.

That answer was not unusual; the count for the whole month is further down this page.

Our page made it into the pool every time. It was read. It was attached. It was not chosen to say anything about us. That gap, between being retrieved and being used, is the subject of this chapter.

Why does a page that ranks well get left out?

Because ranking answers a different question from the one the answer engine is asking.

A search ranking asks: which pages are most relevant to this query, in order? An answer engine asks: which passages, from which pages, let me write a correct, complete answer to this question and its likely follow-ups? Three mechanics separate the two.

The engine runs more than one search. Google says AI Overviews and AI Mode may use query fan-out: issuing multiple related searches across subtopics and data sources to build one response. The question "best invoicing tool for agencies" may also be searched as "invoicing tool pricing", "invoicing tool with time tracking" and "Xero alternatives". A page that ranks first for the head query but says nothing useful on the sub-questions contributes less than a page that answers one sub-question exactly.

The engine shows more, and more varied, links. Google says its AI features can show a wider and more diverse set of links than a classic web search. Slots that were taken by the top ten are spread across more sources.

The engine uses passages, not pages. A ranking rewards a page as a whole: its links, its relevance, its experience. An answer uses a paragraph. If the paragraph that answers the question is vague, hedged, or depends on the one before it, the engine has nothing to lift, however well the page ranks. The passages chapter shows what liftable looks like.

Why can a lower-ranked page be cited instead?

Because the answer engine is not restricted to the top of the list, and lower-ranked pages that state the answer plainly have more to gain.

The GEO study, the first systematic test of content changes against generative engines, found exactly this asymmetry. Adding citations from credible sources lifted visibility by 115.1% for sources ranked fifth, and lowered it by 30.3% for sources ranked first. Its authors conclude the methods are especially helpful for lower-ranked websites. The study used its own test engine and metrics, so treat the numbers as directional, not as a forecast for your pages. The direction is the point: rank is an advantage, not a lock.

Our own workspace, read over the Attensira MCP on 4 October 2026, shows the same thing from the other side. On "Peec AI alternatives", the source that weighed most in ChatGPT's answers was not a neutral review site. It was peec.ai's own page for agencies, the brand being searched against. Next came alternatives roundups on surfacedby.com, promptwatch.com, geohats.com and rankbits.com. Attensira ranks sources by how much they decided the answer, citation share weighted by position, rather than by raw count, and by that measure a competitor's own page outweighed every list of competitors to it.

Does ranking still matter at all?

Yes, more than many GEO pitches admit. Three reasons.

For Google, it is the gate. Google says a page must be indexed and eligible to show with a snippet to be a supporting link in AI Overviews or AI Mode. It also says there are no additional requirements. If you are not in the index with a snippet, you are not in the pool.

For other engines, retrieval still starts with search. The assistants that cite the web search an index before they write. Being findable for the question and its sub-questions is what puts you in that first pool.

AI Overviews often do not appear. Google says they are shown only when its systems judge them additive to classic Search, and often do not trigger. For many queries, the classic ranking is still the whole result.

So the order is: rank well enough to be in the pool, then give the engine something worth taking out of it.

What makes a page get chosen once it is in the pool?

Four properties, most of which are editing rather than new content.

1. The answer is stated, early, in a passage that stands alone. Name the thing, answer the question in the first sentence, and do not depend on a heading or the paragraph above for meaning.
2. Specifics an engine can reuse. A number with its unit and date, a named method, a direct comparison. Vague claims give the engine nothing it can attribute.
3. Coverage of the sub-questions. Pricing, who it is for, how it compares, what it does not do. These are the fan-out searches, and a page that answers several of them is retrieved for several of them.
4. Your name next to the claim. Our pricing page fails here: it lists plans and prices, so an engine summarising "what AI visibility tools cost" can lift a price range from it without the sentence carrying our name. A passage such as "Attensira, an AI-native SEO agency run by software, starts at its Starter plan" gives the engine a sentence that cannot be used without naming us.

How can you tell which problem you have?

Read the answers, not the rankings. For each question that matters, look at three things over at least a few weeks of answers.

• Your page is never among the sources

  What it means

  Not retrieved: an access, indexing or relevance problem

  What to work on

  Crawlability, rendering, and whether the page targets the question at all

• Your page is among the sources, your name is not in the answer

  What it means

  Retrieved and used as background

  What to work on

  The passage: state the answer with your name in it

• Your name is in the answer, your page is not a source

  What it means

  Named from memory or from other people's pages

  What to work on

  Which third-party pages carry your name, and whether they describe you correctly

• Both

  What it means

  You are winning this question

  What to work on

  Keep the page current and watch for a competitor's roundup appearing above it

Our pricing question sits in the second row. Our three branded questions sit in the last row on both engines. Most of our category questions sit in the first: our pages are simply not among the sources.

Attensira's deciding-sources view does this for every tracked question per engine: which pages decided the answer, whether you are on each, and the next move. You can do a version by hand with a spreadsheet and a few dozen saved answers.

How should an SEO team change its routine?

Keep the routine; add a second loop beside it.

The first loop is classic SEO: keyword research, rankings, technical health, links. It still decides whether you are in the pool, and nothing here replaces it.

The second loop starts from questions, not keywords. Take the questions buyers ask an assistant, read the answers each engine gives, and list the pages those answers lean on. For each question, sort yourself into one of the four rows above. Then pick the cheapest move per row: an access fix, a passage rewrite, a page that answers a sub-question nobody answers well, or a placement on a third-party page that already decides the answer.

The loops share a lot of work. A page rewritten so its key passage stands alone tends to earn a better snippet in classic search too. What they do not share is the scoreboard. Rankings tell you about the first loop. Only the answers tell you about the second.

What should you stop measuring?

Two things that look like progress and are not.

Rank tracking as a proxy for AI visibility. A page can hold position one while the AI answer above it names your competitor. Keep rank tracking for classic search; measure AI answers directly.

Citation position as a rank. The place of your link in an engine's list of sources is not a ranking of your brand and is not comparable with a search position. It is a weak ordering signal at best. Report whether you were named, whether you were linked, and who was named instead, each with its count. The measuring chapter covers how.

The next chapter, passages that survive extraction, is about the one property you control completely: what your page says, and how it says it.