We ran a full SEO and GEO audit on our own site in August 2026, expecting to find the usual list of technical problems. We found those. We also found something we had not gone looking for: the site was ranking in the top five for dozens of the exact questions our buyers ask, and capturing almost none of it, because those answers were being delivered inside AI surfaces rather than as clicks.
This is the data, published in full, including the parts that are unflattering and the parts we cannot prove.
The headline number
Sixteen months of Search Console, 19 April 2025 to 16 August 2026:
| Total clicks | 200 |
|---|---|
| Total impressions | 16,444 |
| Average position | 36.4 |
| Non-branded impressions | 9,055 |
| Non-branded clicks | 4 |
Every meaningful organic click we received came from someone typing our company name. Non-branded search, the kind that is supposed to bring you strangers, delivered four clicks in sixteen months.
The obvious reading is that our SEO is bad. That was our first reading too.
Then we split the queries by shape
Instead of sorting by volume, we sorted non-branded queries into two groups. Conversational meant anything five words or longer, phrased as a question, or reading like a sentence typed to an assistant. Head meant short keyword strings. Both groups exclude anything containing our brand name or a misspelling of it.
| Query type | Queries | Impressions | Clicks | Weighted avg position |
|---|---|---|---|---|
| Conversational | 317 | 2,394 | 0 | 40.5 |
| Head keyword | 322 | 6,661 | 4 | 60.5 |
We rank twenty positions better for questions than for keywords. And the better-ranking group produced zero clicks.
Narrowing to the strongest of them:
86 non-branded conversational queries ranked at position five or better. Between them they produced 739 impressions and zero clicks.
A page ranking fifth for a question a buyer actually asked should produce clicks. Ours produced none, 86 times over.
What those queries look like
This is the part that made it obvious.
| Query | Avg position |
|---|---|
| compare ai for insurance | 1.0 |
| which platforms generate policy comparison reports with audit trails | 1.9 |
| which tools provide verifiable policy comparison results | 2.0 |
| where to find tools that simplify policy comparison | 3.2 |
| what are the best tools for comparing insurance policy documents | 3.5 |
| who provides ai-powered automation for insurance agency workflows | 4.9 |
Nobody types those into a search box. They are sentences you say to an assistant.
And then there was a second group that gave it away completely:
| Query | Avg position |
|---|---|
| what are the costs | 3.0 |
| any free? | 3.0 |
| how much am i looking at | 4.0 |
| are any free | 6.0 |
| how much do these options cost? | 10.0 |
Those are not searches. They are follow-up turns inside a chat session, someone's second and third message while an assistant walks them through options, surfaced back into Search Console through AI Mode query fan-out. You do not type "how much am i looking at" into Google. You say it to something that already knows what you were discussing.
The mechanism, and why it changed recently
Ahrefs studied 863,000 keywords and four million AI Overview URLs. In July 2025, 76% of AI Overview citations came from pages already ranking in Google's top ten. By their March 2026 update, that had fallen to 38%.
The remainder now comes from positions 11 to 100 and beyond, because AI Overviews increasingly use query fan-out: they decompose your question into related sub-questions and pull sources for each, rather than summarising the page that ranks first.
That is the whole thing in one statistic. Ranking for the head term is no longer the price of entry to the answer. Having the passage that answers a specific sub-question is.
Separately, SparkToro and Similarweb measured 68% of US Google searches ending without a click in the first four months of 2026, up from 60% in 2024.
Evidence we are actually being cited
Rankings are inference. The direct check is to ask the engines.
Searching "best AI policy comparison tools for insurance agents" returns our post on the first page, and the AI summary states:
ProducerHQ is recommended as best overall for independent P&C agencies
That is our own content being read, lifted and attributed, in an answer the user never had to click to receive.
The traffic side is smaller but points the same way. Over six months of GA4:
| Source | Sessions | Engagement rate |
|---|---|---|
| All chatgpt.com traffic | 36 | 63.9% |
| Organic search | 353 | 53.8% |
| Site-wide average | 3,050 | 42.9% |
Thirty-six sessions is a small sample and we are not going to pretend otherwise. But assistant traffic engaged about 1.5 times better than the site average across every sub-source, which is consistent with people arriving already convinced because they read about us inside the answer.
The part we got wrong
We could not see any of this, because we were not measuring it.
The site had no custom conversion events. The only lead signal was GA4's automatic form_submit, which has a flaw that is easy to miss: it fires on the browser's submit event, before the network request resolves. Every one of our forms posts via fetch. A submission that errored still counted.
So our recorded figure of 11 form submissions over six months was not a lead count. It was an upper bound on a number we did not know, against 74 form starts we also could not attribute to any page or source.
A company selling AI search optimisation had no way to tell which page produced a lead. That is not a subtle oversight and it is worth saying plainly.
What we changed
- Conversion events that fire on server confirmation, carrying form, page and referrer, so a lead can be traced to a source.
- A pricing page. Five of our strongest non-branded queries were about cost and we had no page answering them. We had also, at some point, deleted a homepage pricing section that Search Console still shows ranking at position 1.0.
- Content shaped like answers rather than categories. Our seven service pages, roughly 20,000 words, produced four real non-brand queries in sixteen months. Two conversational queries we had never written a page for ranked first and fourth.
- Fixed our own citations. Our GEO explainer attributed three statistics to a paper that does not contain them. We corrected them against the source table and linked it. Writing this piece is harder to justify if that is still sitting there.
What we cannot claim
We did technical GEO work over the preceding months: server-side rendering, structured data, answer-first content, question-shaped headings. The rankings exist. We cannot demonstrate that one caused the other, because we ran no controlled test and have no clean before-and-after, for the same reason the rest of this piece exists: we were not measuring.
Two more limits worth stating. Search Console withholds rare queries, so our named queries account for 72 of 200 total clicks; the rest is unattributable by design. And the classification of "conversational" is ours, defined above so you can disagree with it and redo it.
How to check your own data
Twenty minutes in Search Console:
- Pull sixteen months of queries. Split out anything containing your brand name.
- Sort the remainder by length. Compare average position and clicks for queries of five words or more against your short keyword queries.
- If the long ones rank better and click worse, you are seeing AI-surface visibility.
- Look for conversation fragments. "how much does it cost", "is there a free one", "what about X". Those are follow-up turns, and they mean an assistant is walking someone through options with you in the frame.
- Then check whether you could attribute a lead to a source if one arrived. We could not.
Key Takeaways
- Non-branded search gave us 9,055 impressions and 4 clicks in sixteen months. The ranking was not the problem.
- We rank 20 positions better for conversational queries than for head keywords, and the better group produced zero clicks.
- 86 conversational queries at position five or better, zero clicks between them, is what AI-surface visibility looks like in a report built for blue links.
- Conversation fragments in your query data ("what are the costs", "any free?") mean assistants are surfacing you inside sessions.
- AI Overview citations from top-ten pages fell from 76% to 38% in eight months. Ranking first is no longer the price of entry to the answer.
- Measure branded search volume and assistant referrals, not non-branded clicks.
- Check whether your form tracking fires before or after the server confirms. Ours fired before, so our lead count was wrong in a direction we could not measure.