Most of what you read about AI visibility is modelled, sampled or guessed. Unsourced sits on something rarer: multi-engine citation data and server-side crawl logs from our own monitoring network. We turn that into original field research, and we publish the exact numbers so you can check them. When the AI-search story moves, you will also find our read on it here, sourced and linked.
Short, sourced takes on the AI-search news that decides whether you get crawled and cited. When something moves, we say what it means and how to see it in your own data.
We put the same 156 questions to the three AI engines that show their sources. On average any two of them cited the same domain about 8% of the time, and 91% of cited sources were unique to a single engine. That is why one AI visibility number hides more than it shows, and why this has to be measured per engine.
Read it →Every AI-visibility tool hands you a number out of 100. Here is exactly how that number gets made, the sampling and blending behind it, and the measured reason one number cannot represent three engines that agree only about 8% of the time.
Read it →14,227 AI-crawler visits across 18 crawlers, checked against operators' own published IP ranges. 56% verified, 10% we simply cannot check — and why Anthropic going from unverifiable to 92% verified changed the picture. 'Out of range' is not 'fake'.
Read the 2026 report →A bot fetching your page and an AI answer citing your page are two different events. Cloudflare's Pay Per Use just built a payment model on the difference, and it changes how you should read every AI crawler that touches your site.
Read it →The EU starts actively enforcing AI transparency on 2 August, including how providers honour copyright opt-outs. But the mandatory template has no way for anyone outside the company to verify a policy was followed. The one place that check happens is your own server logs.
Read it →One phrase doing three jobs. Being shown in an AI answer, being named as the source, and being sent a visitor are different measurements with different tools — and conflating them is what goes wrong at a client renewal.
Read it →The engineering method: query ChatGPT, Perplexity, Gemini and Grok for your real buyer questions, then read your server logs for AI crawlers — with the exact log lines, the user-agent strings to grep for, and how to prove a crawl is genuine.
Read the guide →We verified 9,543 AI-crawler visits against operators' published IP ranges and forward-confirmed reverse DNS. 76.9% were genuine; 5.4% were impostors wearing an AI brand's name — matching an independent 5.7% estimate almost exactly.
Read the report →Perplexity's Comet Plus reportedly puts a $42.5M pool behind the publishers it cites. When being the cited source is the revenue event, ‘we think ChatGPT mentions us’ stops being an answer.
Read it →We put the 100 most common everyday questions to the 7 leading AI assistants and inspected every source they cited. 18% of those sources led to dead or unreachable pages — a look at how checkable AI’s citations really are.
Read the study →Nobody blocks Googlebot, so scrapers wear its name and walk in the front door. As publishers gain legal rights over AI data, a self-reported name will not hold up. Proof of identity will.
Read it →Bing reports AI citations and four new Copilot views. A real milestone, but the metrics are strictly one engine's slice. Here are four things they leave out — and why each one matters.
Read it →An "AI visibility score" is a modelled guess. Here is what evidence looks like instead, and a simple test you can run on any tool — including ours — to tell measuring from modelling.
Read it →Rankings hold, AI citations collapse, and the quarterly rank-tracking report cannot see it. The way through is to stop reporting a number and start reporting evidence — forensic GEO as a retention anchor.
Read it →Your content is the supply chain now, not the traffic source. Optimising harder feeds the machine answering around you more efficiently. The shift is small and total: stop asking how to rank inside the answer, start asking who extracts your work and what you get back.
Read it →Block-everything and allow-everything are both wrong, for the same reason: they treat every bot as identical when they are nothing alike. There are three kinds of AI crawler, not two, and the missing category is the one that matters.
Read it →Most analytics tools trust the User-Agent, but that name is a claim, not a credential. Anyone can put GPTBot in a header in one line of code. The fix is not to trust names harder. It is to verify identity at the network layer.
Read it →Every figure we publish comes from measured data across the sites we monitor — not simulations or illustrative examples. More studies are in progress; methodology and aggregate data are available on request.
See how Unsourced's server-side telemetry compares with the estimate-based approaches:
Unsourced captures which AI assistants cite you, proves which crawlers really fetched your pages, and re-checks after you act — evidence, not a score.
© Unsourced, the evidence layer for AI search. Our data and findings are free to quote and cite — please attribute to Unsourced and link to unsourced.app.