Blog · Third-party sources

What PR means in AI, and what it means in AI search

Published 4 November 2026 · 7 min read

The short answer: Three fields abbreviate to PR near the word AI, and they share nothing but the letters. In machine learning, PR is precision and recall, the pair of numbers behind a precision-recall curve. In software, a PR is a pull request, a proposal to merge a code change. In marketing, PR is public relations, and its job in AI search is narrow: get an independent sentence about you onto a page an engine is able to fetch.

Search Console shows us people arriving at this site on the bare question, with no second word to say which field they are in. So the disambiguation comes before the argument.

Which PR did you mean

Pick the row that matches the sentence you were reading when the question came up.

If the sentence had a model, a classifier, a threshold or an F1 score in it, PR is precision and recall. If it had a branch, a repository, a reviewer or a merge in it, PR is a pull request. If it had a journalist, an outlet, a campaign or a spokesperson in it, PR is public relations. The three never collide inside one sentence, which is why nobody bothers to spell it out and why the search box keeps filling up with the question.

The rest of this piece spends two short sections on the first two, then stays with the third, because that is the one with an answer-engine problem attached to it.

PR as precision and recall

In classification, precision is the share of the items you flagged that deserved to be flagged, and recall is the share of the items that deserved flagging that you actually caught. The scikit-learn documentation gives the formulas as precision equal to true positives over true positives plus false positives, and recall as true positives over true positives plus false negatives, and describes precision as a classifier's ability to avoid labelling a negative sample as positive. Checked on 4 November 2026.

Plot one against the other as you sweep the decision threshold and you get a precision-recall curve, which the same documentation names in full rather than as PR. The abbreviation is spoken more than it is written. It turns up in retrieval work constantly, including the retrieval that sits underneath an AI answer, so an engineer and a marketer can read the same three-letter heading and take away two different subjects.

PR as pull request

GitHub's own documentation defines a pull request as a proposal to merge code changes into a project. Checked on 4 November 2026. Nothing about it concerns visibility, coverage or reputation, and if that is the PR you came for, you are in the wrong building.

It matters here only because the acronym collision is one of the clearer small examples of why entity disambiguation is work. A retrieval system handed the string "PR" with no surrounding context has three plausible expansions and no basis for choosing. Your own pages have the same problem whenever you abbreviate a product name, a plan tier or an internal term and assume the reader will fill the gap.

PR as public relations, and what it is for

The Public Relations Society of America publishes a definition it describes as still standing: public relations as a strategic communication process that builds mutually beneficial relationships between organisations and their publics. The same page records the body's 1982 definition, about an organisation and its publics adapting to each other. Checked on 4 November 2026.

Note what the definition does not mention. Not links, not coverage volume, not domain authority, and not placement counts. Those are the trade's instruments, and the instruments are what changed when answers started being assembled rather than ranked.

What public relations does inside an AI answer

An answer engine assembles a response from passages it retrieved, and it tends to name where each passage came from. For your brand to be described accurately in one, some retrievable document has to contain an accurate description. Public relations is the discipline that gets such a document written by somebody who is not you.

Why does it matter that the writer is not you? Because corroboration is the thing your own site structurally cannot supply. Wikipedia's notability guideline for organisations is the clearest published statement of where the line falls, and it discounts press releases, material substantially based on them, paid and sponsored articles, contributor posts without real editorial oversight, routine announcements such as a funding round or a hire, inclusion in a best-of list, and passing quotations from an organisation's own personnel used as a source. It also says a single significant independent source is almost never sufficient, and that one wire story reprinted across several papers counts once. Checked on 4 November 2026.

Read as a brief rather than as encyclopedia policy, that list disqualifies a large share of what a retainer usually delivers. A trend piece quoting your founder is a company statement with a masthead on it. One line in which the reporter describes your product in their own words is a different kind of object, and it is the object worth paying for.

The two gates nobody checks before pitching

A placement has to clear two gates before it can affect an answer, and both are public.

The first is retrievability. OpenAI documents four user agents, and says plainly that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, while noting they can still appear as navigational links and that a robots.txt change takes around 24 hours to take effect in its systems. Perplexity documents PerplexityBot for search indexing and Perplexity-User for fetches a person triggered. Anthropic documents ClaudeBot for training data, Claude-SearchBot for search quality, and Claude-User for user-triggered retrieval, and says blocking Claude-SearchBot stops its content being indexed for search. Google's AI features guidance says a page has to be indexed and eligible for a snippet, that no extra markup or AI file buys the slot, and that indexing and serving are not guaranteed anyway. All checked on 4 November 2026. Fetch the outlet's robots.txt before you pitch it, not after the piece runs.

The second gate is whether the sentence survives on its own. A passage that reads "the firm is well placed in a growing category" matches no question anybody asks. A passage naming the company, the category and one checkable attribute does. The mechanics of briefing for that are in digital PR when the audience is a model.

What PR cannot do here

Coverage is somebody else's page, and that is the permanent catch. You cannot edit it, you cannot keep it current, and you cannot stop the outlet changing its crawler policy next quarter. We treat that as the difference between rented and owned visibility, and the practical consequence is repair time: a wrong figure on your own domain is a morning's work, and a wrong figure in a trade magazine is a request.

This studio assesses third-party authority as part of an audit. We do not sell placements, and we would be suspicious of anyone who sells both the measurement and the thing being measured. A bought placement is rented visibility with an invoice attached, and it pays for exactly as long as that page keeps ranking and keeps being fetched.

Honest limits

We cannot show you an engine's internal weighting of a press article against anything else, because none of the engines we check publishes one. What is documented is access: which crawler reaches which page, and what happens when it is refused. The independence criteria quoted above are Wikipedia's, written for its own editors, and we use them as a usefully explicit proxy rather than as evidence of how a model scores a source. We have not run a controlled experiment that isolates coverage from everything else a company does in the same quarter, and neither, as far as we can find, has anyone with access to the systems. Our own published study protocol was registered before its run and has no results yet, so it is not evidence for anything today.

FAQ

Does PR mean public relations or precision and recall in an AI context? Both, depending on the room. In a model evaluation discussion it is precision and recall. In a marketing discussion about AI search it is public relations. In a repository it is a pull request.

Is PR the same thing as AEO? No. Answer engine optimisation covers your own pages, your markup, your crawler access and your measurement as well as the third-party layer. Public relations is one input to the third-party layer and does not touch the rest.

Do AI engines treat a press mention differently from a blog post? They document access rules, not preference rules. What is visible from outside is that both have to be fetchable, and that a page nobody else corroborates carries less weight when an answer is assembled from several sources. We go into that in how assistants choose between PR and your blog.

Can I measure whether PR worked in AI answers? Partially. Run a fixed question set across more than one engine, several runs each, before and after the coverage, and record whether you were named, whether the statement was true, and which sources were cited. The cited-source list moves before the mention does.

Is a press release enough? On its own, rarely. A release syndicated across twenty sites is one statement copied twenty times, and Wikipedia's guidance treats a reprinted wire story as a single source for exactly that reason.

Sources

Related reading

Digital PR when the audience is a modelHow assistants choose between your PR and your blogThe third-party sources AI engines keep citingRented visibility vs owned visibility in AI searchWikipedia, and the brand facts engines inheritYouTube, transcripts and video citations