Blog · Third-party sources

How AI assistants choose between your PR coverage and your own blog

Published 4 November 2026 · 10 min read

The short answer: None of the assistants we check publishes a rule preferring earned coverage to your own pages. What is documented is cruder and more useful: a document has to be fetchable by that engine's crawler, it has to contain a passage that fits the question, and a claim repeated by somebody independent is harder for the answer to drop. Your blog usually wins on facts about you. Coverage usually wins on judgements about you.

People ask this question because two suppliers are arguing over the same budget, and each has an incentive to describe the mechanism in their own favour. The mechanism is public in parts and invisible in others, so it is worth being exact about which is which.

How do AI assistants decide whether to use my PR or blog posts as sources?

Think of it as three gates in sequence rather than a single preference score. A document that fails the first gate never reaches the second, which is why so much of this work turns out to be plumbing rather than persuasion.

Gate one is access. Gate two is fit between a passage and the question that was asked. Gate three is whether anything else in the retrieved set says the same thing. Only the first is documented in detail by the engines themselves, and the published research on the other two is thin enough that anybody quoting percentages should be asked where they got them.

Gate one: can that engine's crawler fetch the page at all

This is the part where the answer is written down, and it decides more outcomes than the content does.

OpenAI documents four user agents and separates their jobs: OAI-SearchBot surfaces sites in ChatGPT's search features, GPTBot gathers training data, ChatGPT-User fetches when a person asks, and OAI-AdsBot checks ad landing pages. The consequence of refusing the first is stated outright: sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they may still appear as navigational links, and a robots.txt edit takes roughly 24 hours to register. Perplexity splits the same way, PerplexityBot for the search index and Perplexity-User for a fetch a person triggered, and notes that the user-triggered fetcher generally ignores robots.txt. Anthropic documents ClaudeBot for training, Claude-SearchBot for search quality, and Claude-User for user requests, and says disabling Claude-SearchBot stops its content being indexed for search. Google's AI features page requires only that a page be indexed and eligible for a snippet, adds that no machine-readable AI file or special markup exists to buy the slot, and says indexing and serving are not guaranteed regardless. All four checked on 4 November 2026.

Now apply that asymmetrically. You control your own robots.txt, your own server, your own paywall, your own Cloudflare rules. You control none of those on the magazine that covered you. A feature in a title that blocks the relevant search crawler is a feature that cannot be retrieved, however well written the sentence about you is. The vetting step is a two-minute fetch of the outlet's robots.txt, and most pitch lists have never had it run on them.

Gate two: which passage fits the question

An answer is assembled from passages, and passages are scored against the question, not against the brand.

That favours different documents for different questions. Asked what your pricing is, a retriever wants a page that states the number plainly, and your own pricing page is usually the densest statement of it in existence. Asked which vendor suits a mid-market team in a regulated sector, it wants a document that compares several options, and your blog post arguing for yourself satisfies a fraction of that question while a roundup satisfies most of it. The underlying economics are in the third-party sources AI engines keep citing.

Google describes its AI features as issuing several related searches across subtopics and data sources before composing a response, so one question can pull candidates for half a dozen sub-questions at once. Checked 4 November 2026. A single document rarely wins all of them.

The published optimisation work points the same way at the passage level rather than the domain level. The KDD 2024 paper introducing generative engine optimisation, arXiv preprint 2311.09735, reports visibility gains of up to 40% in generative engine responses from content-side changes, with effectiveness varying by domain. That is a benchmark result on GEO-bench rather than a promise about your category, and it is about how a passage is written, not about who published it.

Gate three: does anything corroborate it

Here the evidence gets thinner and more honest people hedge.

A claim that appears only on your own domain is a claim with one witness. Wikipedia's notability guideline for organisations, the most explicit public statement of what independence means, discounts press releases and material built on them, paid and sponsored articles, contributor posts, routine announcements, best-of listings and passing quotations from a company's own staff, and says one significant independent source is almost never sufficient. Checked 4 November 2026. That is a guideline for encyclopedia editors and not a retrieval algorithm, and we use it as a proxy because it is written down and the algorithms are not.

What that implies in practice is a division of labour. Facts about you, your pricing, your founding year, your feature list, your support hours, belong on your own site in unambiguous language, because you are the primary source and you can keep them current. Judgements about you, that the product suits a particular buyer, that the onboarding is quick, that the category has a real problem, land better when somebody else wrote them, because you cannot be your own corroboration.

The syndication trap, measured

One published study makes this concrete, and it is uncomfortable reading for anyone counting placements.

The Tow Center for Digital Journalism tested eight generative search tools against ten articles from each of twenty publishers, 1,600 queries in total, and published the results on 6 March 2025. Collectively the tools answered more than 60% of queries incorrectly when asked to identify the article, publisher, date and URL behind an excerpt they had been shown. Perplexity got 37% wrong and Grok 3 got 94% wrong. The study also found tools pointing at syndicated copies on aggregator platforms rather than at the original, including one case where ChatGPT cited a syndicated version of an article from a publisher whose crawler policy blocked it. The authors ran one query per excerpt and caution against extrapolating. Checked 4 November 2026.

Read that as a sourcing lesson rather than a league table. Your placement may be attributed to the aggregator that reprinted it, to a different outlet, or to nobody. A release pushed across twenty sites multiplies the chance that whatever gets cited is not the copy you cared about, which is the practical case for fewer, more distinctive briefings and more of your own clearly attributed pages. The separate automated evaluation in arXiv preprint 2410.22349, published October 2024, found inaccurate citation to be a common limitation across the three answer engines it measured, so a named source attached to a sentence is not proof the sentence was checked.

What the source mix actually looks like

One population-scale measurement exists for one surface, and it is worth holding onto because so little else is measured at all.

Pew Research Center took the March 2025 browsing of 900 United States adults, 68,879 Google searches, of which 12,593 returned an AI summary, and recollected the result pages between 7 and 17 April 2025. News sites supplied about 5% of the links inside those summaries, the same share they held in the standard results below. Wikipedia, YouTube and Reddit together supplied about 15%. Government domains supplied 6% against 2% in ordinary results. Checked 4 November 2026.

Press, in other words, is a minority of the citation pool on the one surface anybody has measured properly, and reference and community material is a larger one. That is Google's summary layer in one country in one month, not a law about assistants, but it is a reason to stop treating a press hit as the default instrument.

A test you can run this week

Stop arguing about the mechanism and look at your own citation sets.

Take fifteen questions a real buyer asks. Run each at least three times on more than one engine, because answers move between identical runs. Record, for every run, whether you were named, whether the statement was accurate, and the full list of cited URLs. Then sort those URLs into three buckets: your own domain, earned coverage, and third-party pages you did not place. The ratio is the answer for your category, and it is usually nothing like the ratio the marketing plan assumes. Ten free checks in the same spirit are in our self-check.

Run it again after a campaign lands. The cited-source list tends to shift before the brand mention does, which makes it the earlier signal, and what a model update does to your visibility explains why you should re-run the control set at the same time.

Where we stand on buying the coverage

We assess third-party authority as part of an audit, and we do not sell placements. Keeping those separate is the only way the assessment means anything.

It also reflects what the asset is. A sentence on somebody else's page pays for as long as that page keeps ranking, keeps being fetched and keeps saying what it said, and none of those are yours to maintain. That is the argument in rented visibility versus owned visibility, and it is why the first fixes in an audit are almost always on pages the client controls.

Honest limits

We cannot see a ranking function, so the ordering of the three gates is an inference from documented access rules, published citation studies and our own audit runs, not a published specification. None of the engines we check states a preference between earned coverage and owned pages, and we have not found a controlled experiment that isolates one from the other with the same brand, the same quarter and the same question set. The Pew figures describe Google AI summaries shown to one panel in one month. The Tow Center figures describe one run per excerpt across eight tools in early 2025, and the tools have changed since. The 40% figure is a benchmark result, not a forecast for your pages. Our own audit evidence comes from a prompt set of about 45 prompts across every AI search engine, run with retrieval on and retrieval off, which is a sample and not a census. The study protocol we published was registered before its run and has produced no results yet, so it cannot settle this either.

FAQ

Will an assistant pick a magazine article over my own page? It depends on the question, not on the publisher. Asked a factual question about you, your own page is often the densest and most current statement. Asked for a comparison or a judgement, a document that covers several options tends to fit better.

Does blocking a training crawler stop me being cited? Those are different crawlers with different jobs. OpenAI, Perplexity and Anthropic each document a search-side agent separately from a training agent, and only the search-side one governs whether your pages can be surfaced in that engine's answers.

If my PR gets syndicated, does each copy count? Treat it as one statement. The Tow Center work found tools citing syndicated reprints instead of originals, so syndication can even redirect the credit away from the outlet you targeted.

Should I stop publishing my own content and spend it on coverage? We would not. Your own pages are the only ones you can correct the same day, and they are where the facts about you have to be unambiguous before any third party can repeat them correctly.

How do I tell which one is working? Log the cited URLs, not just the mentions, and compare the mix before and after. A campaign that changes which documents the engine reaches for is working even in the weeks before the brand name appears more often.

Sources

Related reading

Digital PR when the audience is a modelWhat PR means in AI, and what it means in AI searchThe third-party sources AI engines keep citingNews sites, publishers and the citation economyRented visibility vs owned visibility in AI searchLinkedIn as a cited source in AI answers