AI Agents in Search Console: A 30-Day Study (2026 Data)

By Miki Furman, Founder & CEO | Published July 15, 2026 | Updated August 5, 2026 | Call Force Global is a Caribbean nearshore call center operator publishing its own search data.

Executive summary

Google Search Console reports what "people" searched to find your site. In 2026, a growing share of those "people" are not people. AI assistants browsing on a user's behalf, SEO rank trackers, research agents running boolean sweeps, and at least one machine that fires the same benchmark-style sentences thousands of times all register as impressions in the same report that marketers use to compute CTR and judge their titles. This study takes one real property's raw 30-day query export and separates the layers, with the full classification method published so anyone can run the same split on their own data.

Everything on this page is CFG internal search data from the Search Console property for callforce.global, a B2B services site, across two 30-day windows: June 12 to July 11, 2026 (the original edition) and July 3 to August 1, 2026 (the August update). All numbers were computed from the archived query and page exports with the published script, unchanged between windows. Both aggregate tables are downloadable: current window CSV and prior window CSV.

August 2026 update: the July 3 to August 1 window

On August 5, 2026 this study was rerun against a fresh 30-day export, July 3 to August 1, 2026, using the published classification script with no changes. This is an openly revised study, and the revision is the finding: the machine-pattern share halved, from 39.0% to 19.45%, with no change on our side. Same property, same script, same window length. The two windows overlap by nine days, so they are two snapshots of a moving layer, not independent samples.

The volatility matters more than either number. If a fixed bot-adjustment multiplier had been derived from the June to July window (human-only CTR was 1.62x blended), it would already be wrong: in the current window the same correction is 1.22x. No fixed bot-adjustment multiplier is valid. The human versus machine split has to be recomputed for every reporting window.

Both columns computed by the same published script from archived GSC query and page exports. Windows overlap July 3 to July 11.

Current window results by category (July 3 to August 1, 2026)

Position = impression-weighted average. The bracketed LLM bucket recorded zero rows this window (it held 6 impressions in the prior window). Full table with position bands in the current window CSV.

The position paradox held in the new window. 49.9% of non-human impressions registered at positions 1 to 3 and 99.4% at positions 1 to 10, against 12.9% and 40.3% for human-classified queries, whose impressions sit 42.5% at position 21 or deeper. Of everything the property recorded at positions 1 to 3, 48.3% was still machine-shaped. And the machine layer's entire click performance this window, 3 clicks, came from the one-word query "yes".

The original edition: June 12 to July 11, 2026 (preserved)

The sections below are the original July 15, 2026 edition, preserved as published. All figures in them describe the prior window, June 12 to July 11, 2026.

Key findings (prior window: June 12 to July 11, 2026)

• 39.0% of query-attributed impressions matched explicit non-human query patterns. 130 unique queries across 182 query-page rows accounted for 15,614 of 40,033 impressions in 30 days.
• The non-human layer outranks the human layer. Machine-pattern queries averaged position 3.7, with 68.2% of their impressions at positions 1 to 3. Human-classified queries averaged position 19.7.
• The non-human layer produced exactly one click (0.01% CTR). And that click came from the one-word query "yes" at position 1. Human-classified queries clicked at 0.44%.
• A single machine cluster generated 36.6% of ALL impressions. Eight near-identical Belize-themed benchmark queries produced 14,631 impressions at an average position of 3.6 with zero clicks. The top variant alone logged 5,498 impressions.
• Brand queries carried the clicks. Queries containing the brand name accounted for 79 of the property's 109 total clicks, leaving roughly 30 non-brand clicks in 30 days against 40,033 reported impressions.
• 39.0% is a floor, not a ceiling. The classifier is deliberately conservative. The residual "human" bucket still contains queries like "best nearshore omnichannel call center" (1,534 impressions, position 1.3, zero clicks) that behave like machines but do not match any published pattern.
• The device and country cut isolates the synthetic layer geographically. In the same window's GSC API decomposition, US desktop registered 24,392 non-brand impressions at 0.06% CTR, while human-scale Caribbean segments clicked at 7.3% to roughly 24.5% at comparable positions.

Methodology

Source. One B2B site's Google Search Console property (callforce.global), query and page report exported via the GSC API for the 30 days from June 12 to July 11, 2026: 3,020 query-page rows, 2,440 unique queries, 40,033 impressions, 109 clicks. The export includes query, page, clicks, impressions, CTR, and average position. GSC privacy-filters some rare queries out of query-level reports, so page-level impression totals for the property are higher than the query-attributed total analyzed here; this study covers only impressions that GSC attributes to a visible query string.

August 2026 re-export. The current window was produced the same way: a fresh GSC API export for the 30 days from July 3 to August 1, 2026 (3,883 query-page rows, 3,162 unique queries, 38,932 query-attributed impressions, 158 clicks), classified by the same script with no changes. The same privacy-filtering limitation applies: the property's total for the window was 75,412 impressions against 38,932 query-attributed. The two windows overlap by nine days.

Classification. Every row is assigned to exactly one of nine categories by case-insensitive regex, first match wins, in this precedence order:

The exact executable version of this classifier, including the aggregation code that produced every table on this page, is the reproduction script published next to the dataset: build-ai-search-impressions-study.py. It uses only the Python standard library. Per-category metrics are computed as: impression share = category impressions divided by 40,033; CTR = category clicks divided by category impressions; position = impression-weighted average of GSC's per-row average position.

Update cadence. This study is updated quarterly, with interim reruns when the data is quoted elsewhere: same property, same script, trailing 30-day window, aggregate CSV republished alongside prior editions. Updated August 5, 2026 with the July 3 to August 1 window. Next scheduled edition: October 2026.

Results: impressions, CTR, and position by category (prior window)

All figures: CFG internal search data, June 12 to July 11, 2026.

Position = impression-weighted average. Impression share = share of the 40,033 query-attributed impressions. Full table with row counts and position bands in the downloadable CSV.

The position paradox

The most counterintuitive result is the position distribution. The non-human layer does not lurk on page five; it dominates the top of the results. 68.2% of non-human impressions registered at positions 1 to 3, and 99.6% at positions 1 to 10. Human-classified queries show the opposite shape: only 13.1% of their impressions at positions 1 to 3, and 36.0% at position 21 or deeper.

The mechanism is specificity. A rank tracker querying ("bpo cost" or "outsourcing costs" ...) -site:retellai.com or an agent asking a full benchmark sentence generates a query so narrow that only a handful of pages on the web match it, so whichever page matches ranks first. The practical consequence: on an affected property, "average position improved" can mean "the machines got busier," and a page showing position 2 with zero clicks may have no human problem at all.

Anatomy of a machine cluster

The single largest distortion is one cluster of eight near-identical Belize-themed queries, all benchmark-style sentences stamped with the year, for example "belize healthcare outsourcing hipaa phipa alignment 2026" (5,498 impressions, position 2.0, zero clicks) and "benchmarks for nearshore claims processing in belize 2026" (2,539 impressions, position 3.0, zero clicks). Together: 14,631 impressions, 36.6% of everything the property reported in 30 days, at an average position of 3.6, with zero clicks. No human types the same seven-word benchmark sentence 5,498 times in a month. Any impression-level metric computed on this property without removing that one cluster is measuring the machine, not the market.

The 20 example queries, verbatim (prior window)

These are real query strings that Google Search Console reported as searches for which the property appeared, exactly as exported. Impressions and best position are aggregated across pages for the 30-day window.

Query #20, "yes", is the entire click performance of the non-human layer: one impression, one click, position 1. Queries #13 and #14 are agent prompt fragments leaking into the search box, system instructions included.

Each pattern tells you who is querying. #10 and #12 exclude four AI voice vendor domains by name, which is a competitive-intelligence agent monitoring a vendor set. #16 is an entity-disambiguation sweep for the abbreviation "cfg". #13 and #14 carry their own system prompts. #5 to #8 read like questions typed to an assistant, not to Google, and arrive with tens of impressions each at positions 1 to 3 with zero clicks.

The desktop vs mobile split: where the machines live (prior window)

The query export does not carry device or country, so this cut comes from the same property and window via the GSC API device and country decomposition (non-brand queries). CFG internal search data, June 12 to July 11, 2026:

Non-brand queries only. USA desktop accounts for 92.9% of the impressions across these five segments (24,392 of 26,255).

US desktop behaves like no human population: 24,392 impressions, 15 clicks, 0.06% CTR at position 12. That is where datacenter-based agents, scrapers, and rank trackers register, and it is the segment the machine patterns above concentrate in. Where the same property reaches audiences at human scale and reachable positions, CTR is 7.3% to roughly 24.5%, which is normal to excellent. The snippets are fine. The synthetic layer just buries them in the denominator.

What this means for marketers

Stop reading blended CTR. On this property, blended CTR is 0.27%: a 0.44% human layer averaged with a 0.01% machine layer that happens to hold the best positions. Any title test, snippet rewrite, or CTR anomaly alert evaluated on the blend is evaluating noise. The failure mode is specific: because the machine layer ranks at position 3.7 and never clicks, it drags measured CTR down hardest exactly where you think you rank best.

Run the split before the diagnosis. The method in this study takes about 20 lines of standard-library Python: export the query and page table, pattern-match the query strings, and report human and non-human layers separately. Start from the published regex and add your own site's machine clusters as you find them (look for identical long sentences with implausible impression counts, year stamps, and zero clicks).

Decompose by geography and device. If your property serves identifiable human geographies, the country and device report will isolate the synthetic segment fast. A desktop segment with tens of thousands of impressions and near-zero CTR alongside human segments clicking normally at the same positions is the signature. After this analysis, CFG retired US-desktop aggregate CTR as a KPI entirely and now evaluates snippet performance on US mobile non-brand, bot-regex filtered (baseline: 0.31% CTR at position 23.7).

Do not build content for ghost queries. If query mining feeds your content roadmap, strip the machine layer first, or you will write H2 sections answering questions no human asked. In this sample, 130 of the 2,440 unique "queries" were machine-shaped, including the property's four highest-impression query strings.

Track AI retrievals as their own channel, positively. Many machine-shaped queries are assistants fetching pages for a real user, and that surface is worth winning. CFG tracks it separately (a GA4 "AI Assistant" channel grouping, which grew from 24 to 43 sessions week over week as of July 10, 2026, CFG internal data). Exclude it from CTR; do not ignore it.

Limitations

• Single site, short windows. One B2B property, 30 days at a time. The 39.0% share is not an industry average; it is one honest measurement. Sites with different topical footprints will see different shares. The August 2026 update adds a second, overlapping 30-day window in which the share measured 19.45%, which demonstrates the volatility directly; the two windows overlap by nine days and are not independent samples.
• The classifier is heuristic and conservative. Queries are classified by string shape alone. The residual human bucket demonstrably still contains machine traffic (high-volume position-1 zero-click queries that match no pattern), so the non-human share is a lower bound. A small false-positive risk also exists, for example a real person quoting a phrase.
• No server-log confirmation. User-agent level verification was not available for this window, so classification rests on query-pattern evidence, corroborated by the device and country decomposition.
• GSC impressions are position-weighted and privacy-filtered. An impression registers when the page appears in a loaded results set, deep positions on scrolled desktop SERPs included, and Google withholds rare queries from the query report. Query-attributed totals (40,033 impressions here) therefore undercount the property's page-level totals.
• Aggregate CSV only. The download contains category-level aggregates, not the raw 3,020-row export, because raw GSC exports of this property are operational data. The example queries above are published verbatim since they are public search strings, not user data belonging to this site.

Download the data

The aggregate tables behind every number on this page (nine categories plus totals per window: rows, unique queries, clicks, impressions, impression share, CTR, impression-weighted position, and impression share by position band):

Download the current window CSV (Jul 3 to Aug 1)

Download the prior window CSV (Jun 12 to Jul 11)

The reproduction script that classifies the raw export and generates the CSV: build-ai-search-impressions-study.py (Python 3, standard library only).

How to cite this study

Published under a Creative Commons Attribution 4.0 license. Cite, quote, and republish freely with attribution. Suggested citation:

Frequently asked questions

Related CFG data publications: the Caribbean Nearshore BPO Wage Index 2026 and call center outsourcing statistics 2026.