Number one on Google, and still missing from the AI answer

Why do you rank first on Google and still miss the AI answer?

Usually because you rank first for a different question than the one the buyer asks the model. We sit at the top of Google for "bol.com listing optimaliseren", Dutch for optimising a bol.com listing. That is a how-to keyword: someone wants to know how it's done. A buyer who asks an AI assistant which agency can design their bol.com listing is asking something else. The model searches with the words of that question, and your number-one position plays no part.

Google describes the mechanism itself. AI Overviews and AI Mode may use a technique called query fan-out: several related searches at once, across subtopics and data sources, to build one response. The answer then rests on searches the user never typed.

Our own measurement shows how that plays out. Since 8 September we have asked Google's Gemini model, with search enabled, nine Dutch buying questions every morning. Over seventeen measurement days the model ran more than six hundred searches of its own, and our number-one keyword was not among them once. On 12 and 13 September it left us out of its answer to the bol.com question, while we were still at the top of Google.

Ranking still matters. Your position for one keyword is simply a poor predictor of whether you appear in the answer to a question.

The model searches with the words of the question, not your keyword

Our scan stores, next to each answer, the searches the model ran before it answered. That shows what a model does with a buying question. It takes the words of the question and adds the type of provider. It does not translate the question into whichever keyword you happen to rank for.

Part of this is simply an echo. Our nine questions mention the Netherlands and ask for agencies, so "Nederland" and "bureau" appear in a large share of the model's searches too. What the model leaves out says more. Not one search contained the Dutch words for "how", "tips", "improve" or "optimise". The noun "optimalisatie" showed up a handful of times, and always next to "agency". The model was looking for providers, not explanations.

That is the gap with our number-one page. It is a knowledge base article on how to optimise a bol.com listing. Good for someone who wants to do it themselves, which is why it ranks. For the question of which agency does it for you, it is not the answer.

The fair comparison is our buying keyword "bol listing laten maken", roughly "have a bol listing made". There we sit fourth on Google, sixth on two days. And on the bol.com question the model mentions us often: fifteen of seventeen days, usually first. Fourth on Google, almost always in the answer. First on the how-to keyword, irrelevant to the answer. Position and mention move independently, in both directions.

Nine questions, seventeen days: ranking and mention side by side

In our data, a higher ranking did not go together with being mentioned more often. Next to the AI scan we track our Google positions daily for sixteen keywords. We lined up both series day by day, from 8 to 25 September. The 22 September measurement drops out, because the AI scan returned only errors that morning.

Four questions stand out. On product photography, packaging, brand identity and brand plus store, we are not in Google's measured top for the buying keyword, yet the model mentions us on some of the days. It retrieves pages that are not visible for the buying keyword at all.

The Amazon row shows the other side. In the first seven measurement days we sat at positions five to eight for "a+ content laten maken", and the model mentioned us on five of those days. From 15 September we were at position three, and four on the final day. Over those ten days the model again mentioned us five times. The ranking went up, and the share of days with a mention went down.

Enough to adjust our own planning, too little to turn into a fixed rule. Our ranking scan measures from one location and looks eight to ten positions deep. The AI scan uses one model, Gemini in its Flash version via the API, not the AI Overview you see in Google. These are nine questions in one market, and what we see is correlation, not cause.

Mentioned and listed as a source: always together

In our data the model mentions us only when our domain is also among its sources. The scan records two things separately for each answer: whether our name appears in the text, and whether our domain is among the sources the answer rests on. Across one hundred and fifty-three measurements those two always coincided. Every time the model mentioned us, our domain was among the sources, and the other way round.

That says something about how such an answer comes together. For this kind of question, a model with search names companies it came across while searching, not companies it remembers from training. Without our domain among the sources, we were not mentioned once in seventeen days.

You can see it on 12 and 13 September. On those days the model started with exactly the same first search as on the days before and after. Our domain was not among the sources on those two days, so our name was not in the answer either. Why the model chose other sources those mornings, two data points cannot tell us.

There is a limit to what we can see. The scan stores only the domain of each source, not the page. We know oasecreative.nl was used, not whether the model read our service page or a knowledge base article. We are removing that limit: the scan will also store the full address of each source.

On the condition for taking part at all, Google is clear. A page must be indexed and eligible to be shown in Google Search with a snippet to qualify as a supporting link in AI Overviews or AI Mode. There is no separate route. Indexing is the entry condition. Which searches the model chooses decides whether you take part. Which pages a model then prefers when it has a choice is a second layer, which we covered in why a brand name carries little weight in AI recommendations.

Ahrefs: fewer than half of AI Overview sources are in the top 10

The largest measurement we found points the same way as our small one. Ahrefs analysed more than 860 thousand results pages, with four million URLs cited in Google AI Overviews. Of those URLs, 37.9% also appeared within the first ten blocks of the results for the same search. Ahrefs counts every block separately here: regular results, but also ads, featured snippets and video packs. The rest was split almost evenly: 31.2% sat in blocks 11 to 100 and 31.0% beyond that.

In the first version of the same study, from July 2025, that share was still around 76%. That does not make it a clean halving: Ahrefs improved its method in between and now sees citations it previously missed. Since January 2026, AI Overviews also run on Gemini 3. The author attributes the drop mainly to fan-out: Google appears to take less from the original results and more from the results of the sub-queries. Ahrefs presents that as its interpretation, not as a proven cause.

Take the order of magnitude, not the exact percentage. This is an English-language dataset from a company that sells SEO software, and it covers AI Overviews, not the Gemini model we query. But the outcome fits our series: a large share of the sources do not rank at the top for the user's search. If you only track your rankings, you see only part of the road to the answer.

Four clicks on nearly nineteen hundred impressions

A number-one ranking earns less than it used to, even without an AI answer. GrowthSRC, an SEO agency, looked at Search Console data from more than thirty of its client sites in e-commerce, SaaS, B2B and EdTech. The average click-through rate at position 1 fell from 28% to 19% between 2024 and 2025. At position 2 it fell from 20.83% to 12.60%.

Pew Research followed the browsing of 900 American adults in March 2025. Users who saw an AI summary clicked a regular search result on 8% of visits. Without a summary it was 15% of visits. A link inside the summary itself was clicked on 1% of visits. That measurement predates the broad availability of AI Mode.

Our own picture is starker. In the week of 16 to 22 September our site received nearly nineteen hundred impressions on Google and four clicks. In the Search Console query table, all four are on our own name. For "hoger ranken op bol.com" (ranking higher on bol.com) and "ai blogs schrijven" (writing blogs with AI) we sit around position two, together good for dozens of impressions, and zero clicks. Search Console does not show whether an AI Overview appeared above them, and those keywords are not in our AI scan. So we cannot name the cause per keyword. What is certain: a top ranking that earns no click earned nothing that week.

Measure the question, not the keyword

Start with the question the buyer asks. This is what we do ourselves, and it takes no software.

1. Write the question down the way a buyer asks an assistant. Not "bol listing laten maken", but "which agency in the Netherlands can design my bol.com listing?". Write five, for your most important services or products.
2. Ask those questions every day for a week. Use an assistant with search and note which sources appear. In our series the set of companies mentioned changes a lot from day to day, so one measurement says little.
3. See what kind of source wins. Providers' own pages, or third-party lists and directories? For us, own pages win on the bol.com question and lists win on the Shopify question. That decides your route.
4. Put your page next to the question. Does it answer who can deliver, for whom and where? A how-to article rarely does, however well it ranks. A service page or product page should.
5. Count per week, not per day. Per question: on how many of the last seven days were you mentioned? To see the traffic that follows, read measuring AI referrals in GA4.

For an online store the same logic applies, with product questions instead of questions about agencies. The Amazon question in our scan comes closest, and there too the mention did not follow the ranking. A buyer asks about a specification such as size, material or delivery time. If that is not on your product page as readable text, a high ranking on the category name helps little. What a model can read from a product page is covered in getting your store and bol listing ready for AI search results. Part of product recommendations also runs through feeds and catalogues, outside the search results; see agentic search and your store traffic. The basics of GEO are in GEO for webshops.

Next to every how-to article, a page that answers the buying question

We are shifting time from rankings to questions. Three things change.

Our number-one articles stay as they are. They do what they should for people who want to do the work themselves, and an indexed, well-ranking page is the condition for everything above. We are not turning them into sales pages.

Next to every strong how-to article there should be a page that answers the buying question: what we make, for whom, from where. For bol.com listings we have that service page, and on that question we are in the answer fifteen of seventeen days. Whether the model uses that exact page, we will only know once the scan stores the address of each source. On brand books and Shopify we are at zero. For brand books we have no page of our own for the question. For Shopify, third-party lists win, and a better page of our own demonstrably helps little there.

And we report differently. In every SEO engagement a client gets, next to rankings, a weekly figure per buying question: on how many of the seven days were you mentioned, and which source came with it. That figure moves more slowly than a ranking and says more honestly what happens in the answer.

What we do not promise is a place in the answer. The model changes from day to day and nobody steers that. What we can do is make sure that for every buying question there is a page of yours worth retrieving. If you want to know which questions your customers ask and what a model does with them, we can go through your five most important buying questions in a short call.