What a study of ChatGPT snippets found near the page heading

Resoneo’s analysis of ChatGPT retrieval offers a useful reason to inspect the top of a page: some retrieved snippets were short and often included the H1. That is a finding from a captured sample, not a permanent specification for how ChatGPT reads every website.

The retrieval paths in the study

A retrieval pipeline is the process that fetches web results for an answer. The study identified several routes: some retrieved Google results, while another queried OpenAI’s own web index, labelled labrador in the captured responses.

An index stores information about pages, such as their titles, addresses and snippets. In this study, the researchers could identify which retrieval route supplied a result. That made it possible to compare snippet formats across routes.

Until 21 July, ChatGPT’s server responses labelled which pipeline fetched each result. Resoneo read those labels across 1,249 conversations captured in July and published its analysis. Once the labels disappeared, researchers had less direct visibility into the retrieval path.

Independent analyst Suganthan Mohanadasan also investigated source selection in his source-selection census, correcting an earlier assumption when the July data became available.

Licensed and unlicensed sites appeared in the index

The earlier theory was that OpenAI’s own index only held content from publishers it had paid — Le Monde, WSJ, Condé Nast and the rest — and everyone else was reached by scraping Google.

The study found evidence against that assumption. Resoneo found hundreds of outlets with no OpenAI agreement being served from labrador with the same snippet format, the same length, and the same freshness as the licensed partners. Search Engine Journal’s write-up of the findings leads on exactly that point.

There is a contractual difference, but it sits somewhere else. Resoneo found a wordlim value attached to results — a cap on how many words of your page ChatGPT may quote in its answer. Known licensees came in at 100 words, a handful of outlets at 25, and the default for the rest of the web was 200. The unlicensed default is the most generous of the three.

Resoneo also argued that these results did not originate from Bing. One clue was title length: a quarter of labrador’s stored titles exceeded the 75 characters Bing displays. This is part of the researchers’ interpretation of the captured responses, not a published OpenAI specification.

The short snippets near the H1

In the studied results, labrador snippets were cut at 202 characters and drew from page text rather than the meta description. (The Google-scraping pipe still uses meta descriptions roughly a third of the time, so the tag is not dead. It’s just irrelevant to this one index.)

Those 202 characters break into three parts, at Resoneo’s median measurements: about 7 characters of whatever sits above your H1, then the H1 itself at around 51, then roughly 146 characters of body text. SEJ reports the H1 landed inside the snippet in 83.6% of the cases reviewed — 387 of 463.

The H1 was a frequent anchor in this sample, but text before it also occupied part of the snippet. A label above the heading, such as “GUIDES” or “CASE STUDY”, used about 18 characters in Resoneo’s data. Descriptive alt text from an image above the heading could use 50.

Keep that introductory material useful. These measurements are not a reason to remove descriptive alt text needed for accessibility.

Then there’s the plainer failure: Resoneo found roughly one page in seven has no H1 at all. If there’s no heading, the 202-character window starts wherever the visible text starts, which might be your cookie banner copy or a breadcrumb trail.

What to check

• Does the page have exactly one H1, and does it say what the page is about in plain words?
• What sits above the H1? Remove redundant labels where they add no value, while keeping meaningful navigation and accessible image descriptions.
• Read the first sentence after the H1 on its own. Does it stand up as a summary, or does it begin with “In this article we’ll look at…”?
• Don’t rely on your meta description to carry the page’s meaning into AI answers. Make the visible opening useful on its own too.