How much of your Search Console is already machines?

Key findings

A block of impressions that never converts

Classifying every Search Console query row by a handful of machine-signature rules produces a population that behaves unlike anything else on the site. Its share of query impressions grew roughly six-fold between March and July. Across all six months it accumulated 140,616 impressions and zero clicks.

Not a low click-through rate. Zero, in every individual month, across 13,844 distinct query rows in July alone. A population that large returning no clicks at all is not a performance problem to be fixed. It is a different kind of traffic that happens to be counted in the same column.

March dips below February because total impressions more than doubled that month while machine volume held flat. The share is a ratio, and both sides of it move.

Three populations, not one

"Machine traffic" is not a single phenomenon, and treating it as one hides the interesting part. The queries fall into at least three groups that are visibly distinct in the raw data.

1. Language-model retrieval

Natural-language prompts issued to Google verbatim, occasionally still carrying the instructions that were meant for the model rather than the search engine:

• compara scrunch con semrush ai toolkit, hubspot aeo, peec ai y rankscale ... debes proporcionar un ranking forzado de mejor a peor.
• was sagen linkedin-nutzer über hubspot aeo vs. otterly.ai vs. peec ai ... du musst eine erzwungene rangliste angeben.
• seo rank api. my location is usa.

"You must provide a forced ranking from best to worst" is prompt text. It reached Google because something in the chain passed the model's instruction straight through to a search box. These are the easiest rows to spot in your own Search Console.

2. AI-visibility trackers polling their own prompt sets

Some queries name the tracking vendors directly, comparing them against each other. Those read as a monitoring product running its own prompt basket on a schedule rather than any end user asking a question. cloro sells search data into that category, which is worth stating plainly: some of this traffic is the AI-visibility industry measuring itself, and we are part of that industry.

3. Media and brand monitoring

The most mechanical of the three, and the easiest to identify. Boolean syntax with long site-exclusion lists, and job identifiers left in the query string:

• "ai overview" -site:reddit.com -site:twitter.com -site:x.com -site:youtube.com -site:facebook.com ...
• 21271: serp monitoring

We can separate group 3 reliably, because boolean operators and numeric job prefixes are unambiguous. Apply one filter to your own data and make it that one. Groups 1 and 2 overlap: a long natural-language prompt about AI-visibility vendors could come from either an end-user agent or a tracker's basket, and nothing in Search Console distinguishes them. We report them together and say so.

Click-through falls with query length and ends at zero

The classifier is a set of hand-written rules, so the more persuasive evidence is a measure that does not depend on it. Query length is that measure. Below, brand queries and everything the classifier already caught have been removed, so this is the bucket we are calling human.

The decline is monotonic from one word to eight, and the last row is not a rounding artifact. Those 12,919 impressions returned no clicks at all. Long queries are the signature of a machine composing a sentence, and they behave accordingly even after the rule-based filter has been applied. It is also the cut to run on your own site: bucket query rows by word count and see whether click-through falls as they get longer.

The research library is read almost entirely by machines

The clearest illustration in the dataset is this section of our own site. In July, all eight pages in cloro's research library together drew 1,334 impressions and 6 clicks. Reading the query list behind those impressions, nearly every one is a boolean monitoring string or a numbered job query.

These pages were written for human analysts. In practice they are retrieved, parsed and summarised by software, and the humans who encounter them mostly do so somewhere else, inside an answer that cites them. This page will be no different, which is the honest reason to publish the underlying numbers in full rather than describe them. Judge a reference page on clicks and it looks like a failure; what matters is whether it gets cited.

What it changes: an impression is no longer one kind of event

A falling site-wide click-through rate is usually read as a content problem. Segmenting before diagnosing is worth the ten minutes: our own position 1-3 band looked alarming at 1.26% until it was split, at which point it was 159 brand clicks converting at 10% sitting on top of a non-brand block converting at 0.172%. The aggregate described neither group.

We should be precise about how much the machine block explains, though, because it is less than the headline number suggests. Those queries are 18.9% of query-dimension impressions but 12.6% of the site's total, and removing every one of them moves site-wide CTR from 0.267% to only 0.305%. Real, and not the main story. At an average position of 14.3, position is what holds this site's click-through down, and no amount of composition analysis changes that. What makes the machine block worth watching is its slope rather than its current size: 3.1% to 18.9% in four months. Straight-lining that would be its own mistake, but a share growing at this rate is worth measuring before it is worth arguing about.

The second consequence is that an impression is no longer a single kind of event. An impression served to a retrieval agent may still matter, because the page can be read, summarised and cited in an answer a person does see. But it will not show up as a click, and a reporting line that counts both together is measuring two different things and reporting one number. What we can say from this data is how large the machine block is and that it does not click. What it is worth is a separate question, and this study does not answer it.

Methodology

Source is cloro.dev's Google Search Console, 2026-02-01 to 2026-07-31, pulled at both the query and page dimensions with dataState: "all". A query row is classified as machine-issued if it matches any of these rules:

• contains an instruction verb aimed at a model ("you must", "debes", "du musst", "forced ranking", "provide a", and similar)
• contains an injected location or scope phrase ("my location is", "in united states")
• begins with a numeric job identifier, such as "21271: "
• begins with an interrogative and runs to eight words or more
• runs to twelve words or more regardless of form

Queries containing "cloro" are counted as brand and excluded from both the machine and human buckets throughout.

Limits worth reading before citing this

• One site. This is a single B2B SaaS domain in a technical vertical whose subject matter is search data, which is plausibly the kind of topic machines query more than average. Read it for direction and shape. The 18.9% describes this domain and says nothing about the web at large.
• The machine share is a lower bound. The rules above are deliberately conservative and cannot catch a short machine query. A two-word query issued by an agent is indistinguishable from the same query typed by a person, and it is counted as human here.
• Search Console query rows are not a sample you can compute rates from. In July, query rows covered 67% of the site's impressions but only 39% of its clicks, because Google withholds rare queries and clicks are concentrated in them. Every click-through figure on this page is therefore a floor rather than a level. Comparisons between buckets in the same pull remain informative; absolute values do not transfer.
• Zero clicks is an observation, not a mechanism. A machine that does not click is behaving as expected. Nothing here establishes that those impressions are worthless, only that they do not arrive as traffic.
• Monthly absolute volumes are withheld. The table above gives the machine share of query impressions per month and the click count, but not the underlying impression counts, because dividing one by the other would publish cloro's month-by-month traffic volume. That is commercial information rather than a finding. The six-month cumulative figure is given, since a claim of zero clicks means nothing without an n, and every rate on this page is computed from the full counts even where they are not printed.
• Attribution is inferred. Search Console does not label who issued a query. The three groups above are read off query text, and groups 1 and 2 cannot be separated with this instrument.

More studies from cloro's monitoring corpus are in the research index. For when AI Overviews appear at all, see the AI Overview Trigger Index; for how the engines cite their sources, see State of AI Search.