When AI citations don’t send traffic: what the 2026 data shows

A trend piece published in July 2026 by Ahrefs’ Louise Linehan draws on Keywords Explorer data and social media discussion to map where AI search measurement is heading. The sharpest finding: one site owner tracked a 1,900% month-over-month rise in ChatGPT citations to a single page and found the spike moved almost no business metrics.

Table of contents

• What this is about
• What the source says
• What the data shows
• What it does not mean
• What to do
• Numbers at a glance
• FAQ

Key takeaways

• Ahrefs Keywords Explorer data reported by Linehan shows search interest in citation measurement outpacing every other category tracked: “ai search tracking” grew 184% and “ai rank tracking” 175% over the period she examined.
• A 1,900% month-over-month jump in ChatGPT citations to one page produced negligible business impact, per an experiment Wil Reynolds shared publicly — the clearest published illustration of the gap between citation volume and conversion.
• Ahrefs estimates AI Overviews produce roughly a 58% click loss; the company treats this as an internal estimate, not a platform-confirmed figure.
• A study reported by Jan-Willem Bobbink found that applying a standard set of generative engine optimisation tactics to a page lowered its citation rate on GPT-4o-mini relative to the untouched version.
• Agentic traffic surpassed 50% of total internet traffic for the first time in June 2026, according to Cloudflare CEO Matthew Prince — a structural shift that changes who (or what) a page needs to be readable by.

What this is about

The measurement problem in AI search has sharpened. Site owners can now see impression data inside Google Search Console’s new AI performance report, but clicks and queries are absent. The gap between “your page was cited” and “that citation sent a visitor who did something” is where most of the industry’s confusion currently lives. The data Ahrefs published in July 2026 puts numbers on how wide that gap can be.

What the source says

Linehan’s piece covers five trends she observed across search volume data, Reddit threads, and LinkedIn discussion. The two with the clearest bearing on how answer engines source and reward content are the ROI measurement trend and the agent optimisation trend.

On the measurement side, Linehan covers Google’s addition of an AI performance report to Search Console, which surfaces visibility data across AI Overviews and AI Mode. Brodie Clark called it the single most-requested addition Search Console had ever seen. The industry response was largely critical. Simone De Palma, TUI’s Technical SEO Manager, posted a pointed question under Google’s announcement. Why does the report fold AI Overview and AI Mode impressions into one undifferentiated figure? Bing already exposes the individual retrieval-augmented generation queries driving its answers.

The Reynolds experiment is the piece’s most concrete data point. Per that piece, Wil Reynolds noticed a 1,900% month-over-month jump in ChatGPT citations to a single page on his personal site, then measured the downstream effect and found it made little-to-no business impact. That was the moment the industry started asking not “are we cited?” but “what is a citation actually worth?”

On agent traffic, Linehan cites Matthew Prince, Cloudflare CEO, who revealed in June 2026 that agentic traffic — bots, crawlers, and AI agents — had exceeded 50% of total internet traffic for the first time. The implication for page structure is direct. Agents parse structured text reliably and struggle with JavaScript rendering or layout-implied meaning. That is why llms.txt files, markdown page variants, and accessibility tree optimisation all became active discussion topics in the same period.

One finding from Jan-Willem Bobbink deserves careful attention. Bobbink described a study in which a page was put through what he called a typical GEO checklist — adding statistics, quotations, source citations, and an authoritative tone — and tested across three AI engines. On GPT-4o-mini, the baseline page held a 13.3% citation rate; after the optimisation treatment, that figure dropped into the 10.9%–12.2% range. Linehan’s paraphrase of Bobbink’s conclusion: the tactics being widely sold underperformed the untouched control.

Addy Osmani, Director at Google Cloud AI, has loaded “AEO” with a second meaning, repurposing the acronym from Answer Engine Optimisation to Agentic Engine Optimisation. The phrase now generates 80 monthly searches. Suganthan Mohanadasan, AI SEO Researcher at Snippet Digital, pushes back on the renaming exercise, arguing that the protocols themselves are the novel layer — not a relabelling of existing fundamentals.

What the data shows

Search volume figures come from Ahrefs Keywords Explorer. Linehan names the tool and her own employer, so the numbers carry a vendor disclosure. The figures are trend indicators, not independent audits.

• “ai search tracking”: +184% over the period Linehan specifies
• “ai rank tracking”: +175%
• “generative engine optimization”: +997% over 18 months
• “geo vs seo”: +982% over 18 months
• “programmatic seo” (the term itself): +124% over 18 months
• “zero-click search strategy”: climbing (no specific percentage given)
• “agentic engine optimization”: 80 monthly US searches (breakout from near-zero)

The 58% click-loss estimate for AI Overviews is Ahrefs’ own calculation. Linehan labels it “our own estimates” — the source is the same organisation publishing the piece. No methodology detail is provided in the article: sample size, query set, date window, how “click loss” was defined, whether it covers all query types or a subset, and how AI Overview presence was detected are all absent. Per my standard practice here, I am not treating 58% as an established fact. It is an internal vendor estimate and should be read as one.

The Bobbink study is the most structurally interesting data point, and also the most methodologically opaque as reported. Linehan gives the GPT-4o-mini baseline citation rate (13.3%), the post-treatment range (10.9%–12.2%), and the engine tested. What is not reported: the number of prompt runs, the prompt set used, the other two engines tested, the date window, or how citation was defined. The directional finding, GEO tactics did not improve and may have reduced citation rate, is plausible and worth taking seriously, but the reported numbers should not be cited as a confirmed percentage without the full methodology. I am noting this rather than discarding the finding, because the direction matters even if the precise figures are uncertain.

The Reynolds experiment is a single-site observation, not a controlled study. A 1,900% citation increase on one page tells you something real about the disconnect between citation volume and conversion, but it cannot be generalised to a citation rate or a traffic model.

What it does not mean

The natural overreach here is to read these trends as evidence that AI citation work is pointless, or conversely, that it just needs a different tactic. Neither follows from the data.

The Reynolds finding means one site owner found that citation volume did not correlate with conversion on one page over one measurement period. It does not mean citations never drive traffic or revenue. Linehan’s own framing is careful: the question the industry is now asking has moved from tallying mentions to examining what those mentions are actually worth. That is a different claim from “citations are worthless.”

The Bobbink study is the one most likely to be misread. A single study, with methodology gaps as reported, showing that a GEO checklist underperformed an untouched page on one engine is not evidence that all structured content work is counterproductive. It is evidence that the specific tactic set tested, on that engine, in that test, did not help. The finding is useful precisely because it introduces doubt about a practice that was being sold with confidence. Falsifying that doubt would require a larger prompt set, multiple runs, disclosed methodology, and replication on other engines.

Matthew Prince presented the 50% agentic traffic figure as a first-ever milestone for the internet as a whole. Whether that figure holds for a given site depends on the site’s category, audience, and existing bot management. The source acknowledges this directly, and suggests readers check their own bot analytics rather than assume the figure applies.

The absence of click data from Google’s AI performance report is a real limitation, but it does not make the impression data useless. Impressions without clicks cannot answer conversion questions, but they can establish whether a page is appearing in AI surfaces at all, a prerequisite for any further analysis.

What to do

Ordered by effort, lowest first:

1. Find the crocodile mouth in your Search Console data. Linehan describes a specific pattern: pages whose impressions held steady or rose while clicks fell. That divergence is the signal that an AI Overview absorbed the traffic rather than a ranking drop. Identifying those pages takes an afternoon and costs nothing. They are your highest-priority candidates for either content restructuring or channel diversification, depending on what the page is actually for.

2. Separate citation volume from citation value before reporting either number. If you are tracking AI citations with any tool, add a second column: did that citation produce a session? Did that session produce an event? A citation that does not send a visitor is a brand signal, not a traffic channel, and collapsing the two inflates the apparent value of the work. Reynolds’ experiment is the clearest published illustration of why this separation matters, and our own read of 13 months of referral data points the same way.

3. Check which crawlers are actually fetching your pages and at what rate. Agentic traffic exceeding 50% of total requests, per Prince’s disclosure, means the population of non-human readers is now larger than the human one on many sites. Before optimising for agent readability, establish a baseline: which agents are hitting your pages, how often (and what bot-access decisions look like across the web), and whether your current page structure (particularly JavaScript-heavy rendering) is likely to cause parsing failures. This is a diagnostic step, not a commitment to any particular optimisation approach.

4. Before running GEO tactics at scale, design a controlled test. The Bobbink finding is a reason for caution, not a reason to stop experimenting. It lands next to the other GEO claims that thin out under testing. A minimal valid test would compare a treated page against an untouched equivalent page on similar queries, across multiple prompt runs, on at least two engines, with the citation rate recorded each time. Single-run screenshots are not evidence. If the treated page consistently outperforms across runs and engines, you have something. If it does not, you have saved yourself the effort of rolling the changes out further.

Numbers at a glance

FAQ