ChatGPT does not run on Bing. It runs on its own index, and in the mode nine out of ten people use it reads only your page title and about 200 characters starting at your headline.

Google shipped an embeddable button that a publisher adds to its own pages in two lines of HTML, letting a reader mark that site as a preferred source in one tap. Google says more than 600,000 unique sources have been selected. The part that matters is in the documentation, not the announcement. Google's exact wording is that when a user selects a site as a preferred source, its content is more likely to appear in Top Stories with a "preferred" badge, and that "In AI Mode and AI Overviews, your content can be highlighted with a 'preferred' badge for users who have selected your site as a preferred source." Eligibility is domain and subdomain level only: `example.com` and `code.example.com` qualify, `example.com/blog` does not.

Read the difference carefully before anyone repeats it. Google's documentation describes a badge on content in AI Mode and AI Overviews. Several outlets covering the launch went further and said preferred sources causes Google to show links from that site more frequently in AI-driven search experiences. Frequency is the stronger claim and Google's own page does not make it.

RESONEO captured 1,249 ChatGPT answers containing retrieval, 88,000 search results, 26,900 distinct pages and 6,400 distinct domains, using six instruments together: fingerprinted test pages, a Chrome extension recording the server stream, multiple free and paid accounts across countries, feature flag tracking, server log analysis, and API instrumentation. Until 21 July the stream carried a `result_source` field naming the pipeline behind every result, with four values: `labrador` (OpenAI's own index), `bright` and `oxylabs` (Google scrapes) and `serp`.

The findings that matter. ChatGPT grounds answers in three layers: a discovery index that finds pages, a reading cache holding full copies of pages already fetched and shared across users, and a small set of pages opened live. Pages that get opened are cited 74% of the time; pages merely retrieved are cited 7%, a tenfold gap. In free instant mode the model is given only the URL, the full untruncated title, and a stored snippet of roughly 200 characters anchored at the H1, with the H1 itself present in 83.6% of pages that had one. Meta descriptions are not used. Past four megabytes a page is not read at all, with no partial read and no truncation. Titles in OpenAI's own index run to about 202 characters and are never truncated, against roughly 60 characters on the Google scrapes. And the index is genuinely its own: about one in three of its URLs appear on Google's first page, against roughly one in twenty on Bing.

Caveats that have to travel. This is one agency's instrumentation of a stream OpenAI has not documented, not an OpenAI disclosure, and none of it is confirmed by OpenAI. The authors say outright that they cannot explain why snippet-less results are cited at double the rate, and that `result_source` disappeared on 21 July, so nobody can re-run the attribution half. Secondary coverage of the Bing overlap quotes a different figure (1.5% appearing in Bing's top 20) from the primary's one in twenty on Google's first-page comparison, so quote the primary and do not mix the two. Separately, Search Engine Journal's 11 August read of the same corpus found that a content licensing deal did not change how a page was stored or served, same format and same length as an unlicensed page, though it explicitly did not test whether licensed pages get cited more often.

A tool vendor published the aggregate of its own users' first checks: 12,535 unique domains, 398,507 scan results across 1,071 campaigns, six engines (ChatGPT, Gemini, Google AI Mode, Perplexity, Claude, Grok), text mentions only with source-only links excluded. The average score, defined as engines that mentioned the business divided by engines checked, was 8.24%. Pooled share of voice was 19.15%.

Caveats. This is a self-selected sample: people who run a visibility check on a vendor's tool are disproportionately people who already suspect they have a problem, so the true market average is probably higher than 8.24%. There is no per-engine and no per-industry breakdown, no prompt count and no collection window for individual scans, and the publisher sells the tracking software. It is a vendor aggregate, not a study.

A rolling analysis built on the Cloudflare Radar API across 330 cities, with 4,223 robots.txt files parsed in the most recent snapshot. The number worth keeping: the 403 Forbidden rate served to AI bots reached 9.64% in July 2026, against 5.67% a year earlier. Anthropic's crawl-to-refer ratio improved to 1,917:1 from 3,386:1 in June, and OpenAI passed Perplexity at 251:1 against 289:1.

The authors state their own limitation plainly: the robots.txt percentages are a share of directives across Cloudflare's monitored domains, not absolute blocking rates across the internet, ratios swing day to day, and some crawlers misidentify themselves.