Impressions but No Clicks? A Third of Ours Were AI Agents, Not People

The number that started this

Our Search Console property took 18,982 impressions and 27 clicks over 90 days. A 0.14 percent click-through rate.

The usual explanations were all available to us. Average position 17.3 is page two, and page two barely gets clicked. Titles could be weak. Descriptions could be generic. We tested each one, and the interesting part is that most of them turned out to be false for our site.

Then we looked at the queries themselves.

What the queries actually look like

Here are real queries our pages were shown for, copied verbatim out of the Search Console API:

render free instance spin down official 2026
railway pricing trial credits official 2026
cloudflare pages free plan custom domain official 2026
official websites jira productboard aha monday clickup asana linear

No person types those. The last one is not even a question — it is a bare list of eight product names with the words "official websites" in front of it.

These have the fingerprints of a research prompt. An LLM asked to find current, authoritative pricing information emits exactly this shape: the literal word official because the prompt asked for official sources, a trailing year because the prompt asked for current information, and sometimes a flat product list with no question around it at all.

Our first classifier looked for the wrong thing

We had already built a query classifier. It looked for search operator syntax — site:, -site:, inurl:, filetype:, quoted fragments — and reported that only 4.8 percent of impressions were machine-shaped. We believed that number and acted on it.

Operators are what a power user types. They are not what an LLM emits. Not one of the four queries above contains an operator, so the classifier scored every one of them as an ordinary human search.

Rewriting the classifier around the actual signature changed the answer by roughly seven times.

The measurement

Ninety days, one property, query dimension. Each query lands in the first bucket that matches, so the earlier rules are the narrow high-confidence ones.

Agent-shaped floor: 32.8 percent of disclosed impressions.

The single strongest line is official-research: 2,465 impressions across 348 distinct queries, and exactly zero clicks. Not a low rate. A perfect zero, across hundreds of independent queries. A population of humans does not produce that.

The page that convinced us it was not a copy problem

Before measuring the queries, we assumed the fault was ours — bad titles, lazy descriptions. That hypothesis is cheap to test and it mostly failed.

Our worst zero-click page ranks at position 5.3 for the query render free instance spin down official 2026. Its title is:

Render Free Tier Spin-Down 2026: 15-Minute Sleep

That title answers the query exactly. Position five, precise title, accurate description, 1,659 impressions, zero clicks.

We checked the rest of the pattern too. Across our top 22 zero-click pages, only 3 carried a generic boilerplate description. The other 19 were specific and well-formed. Nineteen pages doing it right and taking nothing.

A second instrument agreed independently. In 30 days of product analytics, www.google.com referred 4,256 pageviews to the site and produced zero visitors who ever moved a pointer, scrolled, tapped or clicked — while seven people on other sources did confirm in the same window. The gate works. Google's traffic is qualitatively different, not merely under-counted.

What this does and does not mean

It does not mean agents explain everything. Our ordinary-human bucket converts at 0.28 percent on 3,966 impressions, and at an average position of 17.3 that is unremarkable for page two. That half of the problem is ordinary: we rank where nobody looks.

So there are two problems that need opposite work:

• The agent third cannot be won with copy. The reader is a machine. It reads the result snippet and has no reason to click, no reason to subscribe, and no reason to buy. Time spent rewriting titles and meta descriptions for those impressions cannot pay off.
• The human half needs position, not copy. Page two to page one is a ranking problem, and no amount of title polish substitutes for it.

Conflating the two is what makes a dashboard say "impressions are growing" while nothing downstream ever moves.

Two limits, both cutting against the finding

32.8 percent is a floor, not an estimate. Search Console discloses only a fraction of impressions at the query dimension — 42.1 percent for us over this window — and it specifically withholds rare queries. Long machine-generated strings are rare by construction, so they are disproportionately likely to sit in the withheld half. The true share is higher than what we can show.

Query shape is evidence, not proof. The word "official" in a query is a strong tell. It is not a confession. Nothing here is a measured count of robots; it is a measured count of queries carrying an agent-shaped signature.

And this is one property, on a small site. Take the method, not our percentage.

The part that is actually good news

If a third of your search exposure is agents reading snippets, that exposure is still worth something — it is just a citation channel rather than a traffic channel. The agent is answering someone's question using your page. It will never appear in your analytics.

That changes what is worth building. Machine-readable, well-sourced, dated facts get cited. Marketing pages do not. We responded by publishing our free-tier dataset as JSON — 87 tools, each row carrying the limit that forces a paid plan, the next step's price, the date it was checked and the vendor pages it was checked against — and pointing llms.txt at it:

GET https://www.toolpick.dev/api/v1/free-tier-limits?slug=railway

Same rows as the human page, generated from the same file, so the two cannot disagree.

Run it on your own property

The classifier is a single Python file with no dependencies beyond a Search Console credential. It prints the bucket table above for any property you own, along with the disclosure ratio so you can see how much is being withheld from you.

The rules that matter, if you would rather write your own:

1. the standalone word official — the strongest single tell
2. a trailing year on an otherwise ordinary query
3. eight or more words with no question word — the bare product-list shape
4. search operators — real, but rare, and not what LLMs emit

Check the CTR of each bucket separately before you conclude anything. If one of them returns a hard zero across hundreds of queries, you have found the same thing we did.