Search Console AI Mode Prompts: What 28 Days of Our Own Data Show

Search Console AI Mode Prompts: What 28 Days of Our Own Data Show

On one of our blog posts, 23,477 of 26,660 Search Console impressions (88%) came from queries of 10 words or more. The page earned 1 click in 28 days. If you have looked at your own Search Console AI Mode impressions and wondered why a page "ranks" at position 5 and nobody visits, this study is for you.

We pulled every query on ayautomate.com with 200 or more impressions between September 22 and October 19, 2026, and sorted them by length and page type, then read the wording. Long, question-shaped prompts are a small share of queries, but they cluster on a few pages, they sit at better average positions than real searches, and almost none of them click. Some carry fingerprints that look scripted, like persona preambles or a vendor name dropped into an unrelated question.

Below are the numbers, the method so you can repeat it on your site, and what we changed on our own pages. Every figure comes from our own Search Console data. Where we interpret a number rather than report it, the sentence is labeled Inference.

TL;DR

• Our sample: 383 queries with 200+ impressions, worth 336,728 impressions and 2,601 clicks, out of 1,966,193 site impressions and 17,588 clicks in the same 28 days.
• Queries of 10+ words took 12.3% of impressions and 1 click. Queries of 1 to 3 words converted at 0.89%.
• Long prompts "rank" better than real searches. The impression-weighted average position of 10+ word prompts was 5.9, against 8.6 for 1 to 3 word queries.
• They pile onto listicles: 49.1% of listicle impressions came from 8+ word prompts. Tool profiles: 2.5%. Blog articles: 4.8%.
• 76% of long-prompt impressions were about e-commerce automation.
• A few look scripted: 158-word persona preambles, "do not include location references in your response", an "is it" prefix glued onto prompts we already saw, and a monitoring vendor's name inside e-commerce questions.

What data did we use, and what can't it tell us?

We used the Search Console Performance data for ayautomate.com (web search type, all countries, all devices) for September 22 to October 19, 2026, read through Ahrefs' Search Console integration on October 23, 2026. Ahrefs returns 100 rows per request, so we pulled the site's queries in impression bands (200 to 250, 250 to 300, and so on up to 1,000+) and two pages' full top-100 query lists.

The limits matter more than usual here:

1. Search Console does not tell you which queries came from AI Mode. Google's own announcement of the generative AI performance reports (June 3, 2026, rolled out to all sites by August 31, 2026) says the new view shows impressions by page, country, device and date. It has no query dimension. Google also says this data "is included in the overall performance report", so AI Overviews and AI Mode impressions sit inside the same totals as classic search.
2. We can't see who typed a query. Search Console exposes no user agent or session data, so "this prompt was sent by a tracking script" is always an inference.
3. Our sample covers 17.1% of site impressions. The rest is spread over thousands of queries under 200 impressions plus queries Google anonymizes. Long prompts are, by nature, rare per wording, so this sample probably undercounts them.

Inference: Because Google folds generative AI impressions into the main report, any query-level analysis of "AI Mode traffic" is pattern matching on the query text. That is what this study does, and nothing more.

How much of our Search Console data is long prompts?

Here is the full 383-query sample split by query length.

Two things stand out. First, the click cliff sits at about 8 words: below it, CTR is close to 0.9% whatever the length; above it, 45 queries and 47,024 impressions produced a single click. Second, the long prompts sit higher on average than real searches. A query that ranks at 5.9 and never clicks is not behaving like a classic blue-link result.

Inference: If those 47,024 impressions had clicked like our 1 to 3 word queries (0.89%), they would have sent about 420 visits. They sent 1. Treating them as missed clicks, the way most CTR audits do, overstates what a better title could ever win back.

What do long prompts look like next to real searches?

Long prompts differ from short queries on more than length.

A long prompt surfacing several of our pages at once fits an AI answer that cites multiple sources, or a fan-out of sub-searches behind one prompt. Inference: we can't confirm the mechanism from Search Console, but the pattern is the opposite of a person picking one result.

Which pages attract the prompts?

Prompts are not spread evenly. They gather on "best X tools" listicles.

By topic, 35,754 of the 47,024 long-prompt impressions (76%) were e-commerce automation questions. Content pipeline and marketing operations prompts took another 6,508.

Inference: a listicle that names many vendors and answers "which tool is best for X" is the page type an AI answer would pull from for a "best tool for X" prompt. That would explain the concentration. It would also explain why a vendor-monitoring script, which asks those exact questions, keeps landing on the same pages.

What does a prompt-heavy page look like up close?

We pulled the top 100 queries for our two e-commerce listicles. They show the pattern at full strength.

On the first page, the 10 biggest queries alone account for 82.9% of covered impressions, and every one is a question of 10+ words, such as "which ai ecommerce automation platform gives the most value for the price in usd?" (3,614 impressions, position 5.5, 0 clicks). The short searches for the same page, like "ecommerce automation tools" (position 13.8), add up to about 1,150 impressions at an impression-weighted position of 15.5.

The two pages also share wording. 17 identical prompts of 10+ words appear in both pages' query lists.

Inference: if you judged these pages on position alone, you would think they rank top 5. If you judged them on short queries alone, you would think they rank on page 2. Both are true at once, and the blended Search Console average (6.1 and 10.9) describes neither.

Which fingerprints suggest a tracker, not a person?

Most long prompts read like a plausible question. A few carry text no person types into Google. These are the patterns we found, with exact counts.

Profound is an AI visibility tracking platform. We list it in our AI visibility tools directory. A question about Shopify growth tools that names it unprompted reads like a brand-monitoring prompt set. Inference: we have no evidence about who ran these prompts, and nothing here suggests Profound itself did. Any team tracking that brand, including its customers or competitors, could produce the same queries.

The persona and instruction prompts are the clearest case. They are system-prompt text: a buyer persona, a location rule and a formatting rule, concatenated with the question. Inference: something sent a full prompt template into a Google surface that accepts long natural language input, and Search Console logged the whole string as a query.

Scale is the honest caveat. The fingerprinted queries with the strongest signal (persona, instructions, echoes) add up to fewer than 200 impressions. The bulk of long-prompt volume is clean questions like the value-for-price prompt above. We can't tell from Search Console whether those 3,614 impressions came from a few scripts asking daily or from real shoppers asking in AI Mode.

Are long prompts the only zero-click anomaly?

No, and this keeps the finding in proportion. The three biggest zero-click queries on our site are short: "woodpecker pricing" (40,332 impressions, position 5.8), "woodpecker email automation" (28,515, position 7.3) and "woodpecker review" (18,918, position 2.8). That is 87,765 impressions on one tool profile and zero clicks.

Query length does not explain that. We don't know yet what does, and we are not going to guess in a data study. The method below catches prompt-shaped noise; it does not clean every inflated number in your report.

How do you separate prompts from searches in your own Search Console?

You can repeat this in an afternoon with a Search Console export or the API.

1. Export queries with enough volume. Use 28 days, web search type, and a floor (we used 200 impressions) so each row is stable.
2. Bucket by word count. Split at 1 to 3, 4 to 7, 8 to 9 and 10+. In our data the click cliff sat at 8 words; check where yours sits.
3. Flag question shape. A trailing "?" separated the groups sharply for us (76% of long prompts against 0.6% of short queries).
4. Flag repeats across URLs. Count how many of your URLs each query surfaced. Prompts averaged 2.67 URLs, short queries 1.96.
5. Search for fingerprints. Look for "my main", "do not include", "in your response", "context:", queries starting "is it", brand names that don't belong in the question, and URL parameters like "udm=".
6. Re-score pages without the prompts. Recompute position and CTR from short queries only. A page that looked like "position 6, terrible CTR" may really be "position 15, normal CTR for page 2".
7. Cross-check with the generative AI report. Pages that rank for many prompts should also show impressions in Search Console's generative AI view. That report has no queries, but it confirms which pages appear in AI features at all.

Rule of thumb from our data: if more than half of a page's impressions come from 8+ word questions, stop judging that page on CTR. Judge it on whether AI answers cite it, and on the short queries you actually want to rank for.

What did we change on our own pages?

This analysis came out of our own SEO loop, and it changed the fix. On October 22 and 23, 2026, the loop's first batch looked at seven listicles with high impressions and very low CTR, the pages a classic CTR audit flags as the biggest "missed click" opportunities. On the prompt-heavy ones, rewriting titles would have targeted impressions that do not click. Instead we:

• added short, direct Q&A answers to the exact prompts each page ranks for (for example, which e-commerce automation platform gives the most value for the price), so an AI answer can quote the page;
• rewrote three intros to lead with the answer;
• kept titles unchanged on four pages that had been retitled a day earlier, so each change stays measurable;
• added internal links into the pages with few inbound links.

We will measure the result on November 6 and November 20, 2026, and judge the prompt-heavy pages on short-query clicks and AI citations, not blended CTR. If you want the same treatment for your site, our SEO and GEO content system is the workflow we use, with real search data in and human review on every draft.

For the citation side, see our guides to how to get cited by ChatGPT and Perplexity and GEO for AI search, and our earlier look at zero-click AI search data.

FAQ

Does Search Console show AI Mode queries separately?

No. Google's generative AI performance reports, launched June 3, 2026, show impressions by page, country, device and date, with no query dimension (Google Search Central). Google says the same data stays inside the overall Performance report, so AI Mode queries mix with classic search queries. You can only separate them by patterns in the query text.

Why are my Search Console impressions high but clicks so low?

Long, prompt-shaped queries are one common cause. In our 28-day data, queries of 10+ words were 12.3% of impressions and produced 1 click, while sitting at a better average position (5.9) than short searches (8.6). Check what share of a low-CTR page's impressions come from 8+ word questions before you rewrite its title.

What counts as a long prompt query?

In our data the useful line was 8 words. Below it, CTR held near 0.9%; at 8 words and above it dropped to effectively zero. Long prompts were also phrased as questions (76% ended with "?") and surfaced more of our URLs per query (2.67 against 1.96).

Are these queries from AI visibility tracking tools?

Some show clear signs of automation, such as 158-word persona preambles and "do not include location references in your response". Search Console has no user agent data, so attributing any specific query to a tracking tool is an inference. In our sample the clearly scripted queries were small in volume; most long-prompt impressions were clean questions whose source we can't determine.

Should I filter long prompts out of my SEO reports?

Report them separately rather than delete them. Recompute position and CTR from short queries to see where a page really ranks in classic search, and track prompt-heavy pages on AI citations. Our SEO and GEO content system handles both views.

Which pages get the most AI-style prompts?

Comparison and "best tools" listicles, in our case. 49.1% of our listicle impressions came from 8+ word prompts, against 4.8% for other blog articles and 2.5% for tool profile pages. Pages that answer "which tool is best for X" are the ones those prompts ask about.

---

Sources: ayautomate.com Search Console data, web search type, September 22 to October 19, 2026, via Ahrefs' Search Console integration (pulled October 23, 2026); Google Search Central, Introducing Search Generative AI performance reports in Search Console.