Cloudflare is now selling AI visibility from inside the pipes, and a 50,000-prompt study shows the ad slot and the recommendation are two different things.

Five days old rather than inside the 72-hour window, and not covered here before. During Agents Week (5 to 10 August) Cloudflare put Agent Readiness into the main dashboard and opened early access to an AEO Visibility dashboard. It reports four metrics: citation rate (which platforms pull from your site as a source), mention rate (whether assistants name you without citing you), prominence (how much of the answer is attributed to you and where it sits), and share of voice against competitors on the questions their customers ask. The data comes from querying the models with category prompts and reading the responses, which is the same method every tracker uses, including ours. Coverage today is, in Cloudflare's words, "the leading assistants (today, Anthropic's Claude and OpenAI's GPT)". Access is by request from the Cloudflare dashboard Overview tab. Pricing is not disclosed.

Restaurants only, thousands of them, US and Canada, booking completed without leaving the conversation. It extends the July content partnership that put Yelp reviews and ratings into ChatGPT answers. Yelp's framing: businesses on Yelp are "more likely to be surfaced in key discovery moments where consumers are actively making local decisions on AI platforms". The post mentions requesting quotes from service providers as part of Yelp's wider ecosystem, not as part of this release.

50,006 commercial prompts across 20 US niches, roughly 2,500 per niche, collected 23 July 2026, measured on the free and Go tiers where ads run. Findings, quoted: "ChatGPT shows ads on 25.94% of commercial prompts, close to AI Mode's 29.45%." On relevance: "Across all ad impressions, 14.35% were no more closely related to their own query than an ad paired with a random query from the dataset." And the one that matters most to us: "The advertiser's domain appeared among the cited sources in just 3.63% of placements, while the exact advertised URL appeared in only 0.09%." Category spread is wide. Healthcare drew ads on 28.69% of ChatGPT prompts against 2.64% in AI Mode, more than ten times the rate.

Judge it honestly: one vendor, one collection day, US only, and semantic-similarity scoring for relevance rather than human judgement throughout, though flagged ads were manually reviewed. The sample is large and the method is stated, which is more than most things in this category offer.

Sock-puppet audit: 91 simulated accounts across income and race groups, 335 prompts, more than 127,000 conversations, 3,602 ads from 191 advertisers. Ad exposure fell as account income rose (odds ratio 0.98 per $1,000 of median household income). No significant race effect, and the authors say their sample cannot settle that. Retail and information sectors were about 57% of ads. Ads were visually separated from the model's answer rather than embedded in it. We are citing this through secondary coverage; the arXiv record did not resolve on the identifiers we tried, so treat the figures as reported rather than read.