Cloudflare released the AEO Visibility Dashboard on August 6, 2026, a tool that samples how Anthropic's Claude and OpenAI's GPT models answer buying questions in a site's category, and reports back four numbers describing whether the site is cited, named, or ignored.

The product arrived through Business Wire at 9:00 AM Eastern time on August 6, alongside a technical post on the company blog written by Matthew Conroy and Jack Galilee. It sits inside what Cloudflare calls its Answer Engine Optimization Suite, joining Agent Readiness, an existing tool that checks whether automated systems can reach and parse a site at all. According to the announcement, the dashboard is available in early access only, with businesses requesting entry from the Overview tab of the Cloudflare dashboard. No general availability date was given.

The gap the company describes is one of measurement rather than performance. According to Cloudflare, when a customer asks an AI assistant for a recommendation and receives a name in reply, no signal reaches the brands involved telling them whether they were recommended, mentioned in passing, or left out entirely. Search rankings produced an approximation of that information for two decades. Conversational interfaces produce none.

"Being discoverable used to mean ranking on a page. That's not enough anymore. The agentic era is here, and the moments that matter, like when someone asks an AI for a recommendation and gets an answer back, are happening at scale, invisibly, without most brands knowing if they're in the conversation at all," said Stephanie Cohen, Chief Strategy Officer at Cloudflare, in the announcement.

Four metrics, and what separates them

The dashboard reports four figures, each measuring a different failure mode.

Citation Rate covers the share of answers in a category that cite the site as a source. Prominence measures, when a citation does occur, how much of the answer is actually attributable to that source and how early in the response it lands. According to the announcement, a passing reference at the end of a long answer is not equivalent to a specific recommendation near the top, and the metric exists to put a number on that difference.

Mention Rate tracks how often assistants name a brand in their answer regardless of whether its website is cited as a source. Read against Citation Rate, the pair is designed to separate two problems that look identical from the outside. According to Cloudflare, a brand named far more often than it is cited has an authority problem rather than an awareness problem, a distinction the company argues should change how content budget is allocated.

Share of Voice compares a site's slice of citations against those earned by competitors across the questions customers in that category actually ask.

A fifth figure, described in the blog post rather than the press release, is an Industry Fit score. It measures whether an AI assistant places a site alongside its genuine competitors, derived by mapping which brands consistently surface together in the same category responses.

How the numbers are produced

The methodology is the part of the announcement that carries the most weight for anyone deciding what the figures represent.

According to the blog post, Cloudflare infers a site's industry and category from the site itself, using health and fitness as an example industry and sports apparel as an example category. It then probes what the post calls the leading assistants, which as of publication means Anthropic's Claude and OpenAI's GPT models, with prompts structured to imitate real discovery behaviour: requests for recommendations, product comparisons, and general advice within the category.

Because models rarely answer the same question identically twice, the company routes those prompts through Cloudflare AI Gateway and repeats them multiple times across different models. Responses are then read as a customer would encounter them, meaning the answer text together with the sources each assistant cited. Where scoring requires judgment, Workers AI performs the evaluation on Cloudflare's own infrastructure. According to the post, the pipeline also applies exact text analysis rather than relying on a model to grade its own output.

One architectural choice shapes what the figures can and cannot show. Rather than re-querying models each time a site owner runs a scan, Cloudflare builds a benchmark panel once per industry category and reuses that baseline across every account in the same domain. The company lists three reasons: results load from a stored snapshot instead of waiting on live model calls, aggregating queries at category level avoids duplicate AI calls across thousands of scans, and reusing one corpus makes the Industry Fit comparison possible.

The practical consequence is that the citation and mention figures a site owner sees are drawn from a pre-computed dataset for that category, not from queries issued at the moment of the scan.

A tension worth naming

The press release draws a sharp line between Cloudflare's approach and that of competitors. According to the announcement, most tools attempt to fill the visibility gap by sending test prompts to AI chatbots and sampling the responses, a method it describes as limited in scale and prone to inconsistency unless paired with other data signals. Cloudflare's position, the release states, is different because the company operates at the network layer between AI platforms and the websites they reach, observing crawl and referral activity directly across millions of sites.

The blog post describes the AEO tab doing precisely what the release characterises as the limited method: probing assistants with likely prompts and reading the answers. The network-layer data appears in the product, but in a separate panel.

That panel is AI Operator Activity, which reports real crawl and referral traffic per operator, naming OpenAI and Google among them, along with the errors those operators encounter, including 403 blocked responses and 404 dead links. According to the blog post, the pattern worth attention is the operator that crawls thousands of pages while referring nobody back. The authors state that because requests physically transit Cloudflare's network, the tools measure rather than estimate where possible.

So the suite combines two evidence types with different properties: sampled model output for the answer metrics, and observed network traffic for the operator metrics. The press release compresses both into a single claim about network-layer superiority. Marketers evaluating the product against rival tools would be reading the two documents against each other to establish which number rests on which foundation.

The Agent Readiness half

Agent Readiness predates the visibility dashboard and has now been integrated into the main Cloudflare dashboard. Its technical component, Diagnostics, scans a hostname the way an automated agent reads it: establishing whether it is permitted entry, whether it can discover content, whether it can fetch a clean machine-readable copy, and which interfaces it can call.

Results roll into a single readiness view ranging from "Not Ready" to fully agent-native. Every individual check returns pass, fail, or neutral, accompanied by a note on why it matters and an evidence trail recording the exact request and response observed.

Checks are grouped by the effort required:

  • Quick wins cover a crawler-readable robots.txt, an XML sitemap, AI-crawler rules, and serving clean markdown to agents.
  • Technical groundwork covers Content Signals declaring how content may be used, an API catalog, link headers, and agent login instructions.
  • Advanced integration covers OAuth discovery, MCP and A2A agent cards, a skills index, Web Bot Auth, and WebMCP.
  • Commerce covers x402, described in the post as an extension of the classic HTTP 402 Payment Required status code, alongside ACP, the Universal Commerce Protocol, and AP2.

The commerce group is informational and does not count toward the readiness score, a detail that positions agent payment standards as watched rather than expected. Cloudflare has been building in that area for some time, having partnered with Visa and Mastercard on cryptographic authentication for shopping agents in October 2025 and maintained an x402 proxy template through early 2026.

Each flagged improvement carries a next step. Where a Cloudflare setting resolves the issue, a "Set up in Cloudflare" link points directly at it. Where it does not, a "Copy Agent Prompt" button generates instructions for a coding agent. The checklist overlaps substantially with the Agent Readiness category in Joost de Valk's Website Specification, published on May 30, 2026, which covered 18 topics under that heading and flagged llms.txt as an emerging convention rather than a ratified standard.

Traffic composition behind the launch

The blog post opens with a figure that frames the rest. By Cloudflare's own count, fewer than half of all HTML page requests now originate from a human. The authors note that not every machine request comes from an agent acting for a person, but state that the share is growing.

That measurement is consistent with figures the company has published across 2026. Research produced with ETH Zurich in April documented how crawler behaviour breaks standard caching assumptions, and Cloudflare Radar data from early June put bots at 57.4% of HTML traffic against 42.6% for humans, with training-related crawlers alone at 50.6% of the total. IAB Australia's crawler guidance, published on July 30, cited Cloudflare figures putting automated requests at 57.5% of web-page requests as of June 2026 and described the crossover as the first on record.

The economic asymmetry underneath those percentages has been the company's recurring theme. Its July 1 Attribution Business Insights dashboard showed crawl-to-referral ratios running from 118 crawls per referral at the low end to nearly 50,000 at the high end. On the same date Cloudflare replaced per-crawl charging with a model that pays publishers when content contributes to a generated answer, justified with internal data showing that more than half of crawl traffic from bots it considers legitimate re-fetches unchanged pages. A companion policy sets September 15, 2026 as the date default blocking of Training and Agent crawlers begins on ad-bearing pages for domains newly joining the network.

Where this sits in a crowded category

Answer engine optimization has acquired a commercial software layer over roughly sixteen months, and the tools now differ mainly in where they intervene.

HubSpot launched a monitoring product at 50 dollars per month on April 14, 2026, disclosing at the time that organic traffic for its own customers had fallen 27% year over year, then followed with a free public dashboard called AEO Sensor in May. Typeface approached the problem from content creation rather than monitoring, scoring material for structural problems before publication. Later built a version for creator marketing on May 21, 2026, citing a finding that only 10% of AI search references originate from a brand's own site. Microsoft Advertising had already published a retailer playbook on January 6, 2026 framing the objective as influence rather than traffic.

Demand appears to outrun capability. Adobe research covering more than 500 marketers found that 98% lacked a confident AI search strategy, and separate work published in July found that AI agents could not read pricing on roughly a third of leading B2B sites, with server-side rendering the dominant obstacle.

What distinguishes the Cloudflare entry is placement rather than method. The company sits in front of a substantial share of the web, which gives the operator-activity panel a data source rivals cannot replicate. The answer metrics themselves, however, rest on the same sampling technique the industry has been using since the category formed.

Cloudflare's forward-looking statements attach the usual caveats, noting that the capabilities and effectiveness of the AEO Suite, and the timing of general availability, could differ materially from what the announcement implies, with risks detailed in the Quarterly Report on Form 10-Q filed on May 8, 2026.

Timeline

Summary

Who: Cloudflare, Inc. released the tool. Stephanie Cohen, Chief Strategy Officer, provided the announcement commentary, and Matthew Conroy and Jack Galilee authored the accompanying technical post. The measured assistants are Anthropic's Claude and OpenAI's GPT models. Affected parties are site owners, brand marketers, publishers, and the AI operators whose crawl and referral activity the tool reports.

What: The AEO Visibility Dashboard, an addition to Cloudflare's Answer Engine Optimization Suite. It reports Citation Rate, Mention Rate, Prominence, and Share of Voice for a site within its inferred industry category, adds an Industry Fit score, and displays crawl and referral traffic per AI operator including 403 and 404 error patterns. It joins Agent Readiness, whose Diagnostics component scores sites from "Not Ready" to fully agent-native across four check groups: quick wins, technical groundwork, advanced integration, and commerce, the last of which is informational and excluded from the score.

When: Announced on August 6, 2026 at 9:00 AM Eastern time. Available in early access from that date, with no general availability date disclosed.

Where: Announced from San Francisco through Business Wire and the Cloudflare blog. Access is requested through the Overview tab of the Cloudflare dashboard, where Agent Readiness also sits.

Why: Conversational AI interfaces return recommendations without producing the ranking, impression, or click data that made search performance observable. Cloudflare states that fewer than half of HTML page requests now come from humans, and industry figures put automated web-page requests at 57.5% as of June 2026. The dashboard is positioned to convert model responses and network traffic into figures marketers can compare over time, though the answer metrics themselves rest on sampled model output drawn from a pre-computed category benchmark rather than on the network-layer observation the press release emphasises.