What the evidence says about AI crawler infrastructure in 2026

The web traffic charts most publishers are staring at in 2026 are not anomalies, and they are not recoverable through traditional SEO adjustments. We have spent the past year collecting and adjudicating evidence about what AI-generated search results are actually doing to query economics. The picture is less ambiguous than the industry debate suggests.

Our conclusion: the shift from click-bearing to zero-click search has accelerated, it is measurable across independent data sources, and it compounds an asymmetric penalty on content that does not appear inside AI-generated summaries. The question is no longer whether this is happening. The questions are how fast, how unevenly, and what the evidence says the right response looks like.

Zero-click search crossed a threshold, and the data agrees

Zero-click searches in the U.S. reached 68.01% in the first four months of 2026, up from 60.45% in 2024. That is a 7.56 percentage-point increase confirmed across four independent data sources with a confidence we treat as reliable. Expressed differently: the share of searches generating at least one click fell 22.9% in relative terms over 2024 to 2026, or 9.51 percentage points in absolute terms. These are not projections. They are measured outcomes from infrastructure we track for our audit clients.

The mechanism is not mysterious. Informational queries — the kind that once drove consistent top-of-funnel traffic — now frequently return a Google AI Overview that resolves the query on the results page itself. We have confirmed across four independent sources that informational query click-through rates decline by 34% when a Google AI Overview is present. Visitors who would have clicked do not need to.

Being cited in an AI Overview is not a nice-to-have — it is the only trade worth making

The asymmetry inside a single SERP is stark. Pages cited inside a Google AI Overview see a +5% CTR lift. Non-cited pages in the same search results page see a -25% CTR decline. That is a roughly 30-point spread between citation and non-citation within the same query, confirmed across four independent sources. We have also confirmed that citation inside an overview effectively reverses the traffic penalty: the cited domain does not merely avoid the loss, it gains relative to its prior baseline.

One additional data point we hold with appropriate caution: a single-source analysis suggests the position-1 organic CTR drops from 31.7% to 23.4% when an AI Overview is present. That is a meaningful 8-point drop even for the highest-ranked result in the SERP. We have only one source for this figure, from a lower-authority measurement provider, so we flag it as supported rather than verified. The direction, however, is consistent with every other signal in the dataset.

How many queries actually trigger an AI Overview is still a contested number

Here the evidence is less settled, and we want to be precise about what is and is not established. Current estimates place AI Overview prevalence at roughly 30% of U.S. Google queries as of March 2026, with a prior estimate of roughly 12% in May 2024, implying approximately 2.5x growth over that period. We hold all three figures as contested. Four independent sources do not converge on a single number: the query samples differ, the measurement windows differ, and the methodologies are not reconcilable. What we can say is that the scale is substantial enough to explain the zero-click data above. Similarly, the frequently cited claim that AI Overviews cause an average organic CTR decline of roughly 25% is contested both in magnitude and in how that average is constructed across query types.

The 2% monthly growth figure is wrong — retire it

We want to be direct about one specific claim in circulation: that Google AI Overviews are growing at approximately 2% per month in the U.S. Four independent sources refute this figure. The evidence does not support a smooth, linear monthly growth rate for AI Overview deployment. We do not know the precise rollout cadence — it is neither public nor cleanly inferrable from traffic residuals — but the 2% monthly figure is not a measurement, it is an extrapolation without a credible basis. Any forecasting model built on it is built on nothing.

llms.txt sits at 10% adoption, with the highest-exposure sectors at the back of the line

We analyzed a dataset of 300,000 domains and found that 10.13% had an llms.txt file as of Q1 to Q2 2026. That figure is low, but the sector distribution is the more important story. Finance, healthcare, and legal sectors have the lowest llms.txt adoption rates of any tracked vertical: less than 10% of top-100 domains in each of those sectors had an llms.txt file as of Q1 to Q2 2026. These are sectors with high volumes of informational queries, significant regulatory exposure to AI-generated misrepresentation, and meaningful downstream risk from uncontrolled crawler behavior. They are also, on this evidence, the least instrumented.

Which sectors lead on adoption is a more contested question. The data we have suggests SaaS, publishing, and consumer-facing companies are further ahead, but we hold that ranking at lower confidence and would not present it as settled. What is settled is that the structural vulnerability and the preparation level are inversely correlated.

The practical implication of all of this is not that llms.txt resolves the citation problem — the evidence for its effect on AI Overview inclusion rates does not yet exist in a form we would call verified. The implication is simpler: the measurement infrastructure for AI visibility lags the actual AI infrastructure by a margin large enough to be consequential. We are measuring query economics after the fact, inferring crawler behavior from traffic residuals, and watching the sectors with the most exposure remain the least instrumented. The evidence we have is already clear enough to act on. The evidence we do not yet have is the more urgent problem.

The receipts: every fact, adjudicated

The article above asserts nothing our truth engine did not independently verify. Here is the full ledger — each fact, its verdict, confidence, authority tier, and corroborating sources — so you can check our work. Verdicts are grounded in the fetched content of each source, not a model's priors.

Verified facts

Corroborated and adversarially adjudicated — the evidence bar was met.

Supported facts

Judges agree, but the evidence rests on limited or single-authority sourcing — published with that caveat.

Contested

Evidence is split. Shown with both sides visible; do not treat as settled.

Unverified

Insufficient evidence either way — monitored, not asserted.

Refuted

The truth engine refuted these — shown so they are not repeated as fact.

How we verified this

Of 21 atomic facts: 14 verified, 1 supported, 5 contested, 1 refuted. Each fact was judged by a 3-reviewer adversarial panel prompted to refute it. A fact reaches 'verified' only when corroborated by at least two independent sources or one high-authority source AND the panel agrees. A single low-authority source reaches 'supported' at most — never 'verified'. Sources are scored per-fact: the same article can be right about one fact and wrong about another. This report is regenerated when new research shifts a verdict; the date reflects the last adjudication.