Engines decide how much attribution exists, and you can only compete for your share

Published 29 August 2026. Status: Original concept, first publication.

There’s a paper out of University College London, Waterloo and O’Reilly Media that counted what search-enabled models read against what they credit, and one number in it has been bothering me for a week. Perplexity’s Sonar visits around ten relevant pages to answer a query and cites three or four of them. The other six were relevant, they were read, and they got nothing.

The paper is about whether the web gets paid. The question sitting underneath it, which nobody is asking, is which brand ends up in the three.

Ilan Strauss and his co-authors measured what the engines read against what they credit

The attribution crisis in LLM search results, by Ilan Strauss, Jangho Yang, Tim O’Reilly, Sruly Rosenblat and Isobel Moure, ran across roughly fourteen thousand real conversations and named the shortfall the attribution gap: relevant pages read, minus pages cited. I’m using their term and crediting them for it rather than minting my own, which is the reframe, cite and add I wrote about yesterday, running live.

Their findings are worth having in front of you. Gemini gives no clickable source in ninety-two per cent of its answers, a third of Gemini responses are produced without fetching anything from the web at all, and Gemini and Sonar each leave about three relevant sites uncredited on the average query. Across models answering identical questions, extra citations handed out per extra page read ranged from 0.19 to 0.45, which the authors read as retrieval design rather than technical limits. They’re careful about their own data too: GPT-4o’s small gap is attributed to selective log disclosure rather than to better manners.

Their conclusion is an economic one, about incentive structures and the viability of producing good information, and they recommend standardised telemetry so this can be audited properly. I’d sign all of it.

The paper asks whether the web gets paid, and the question underneath is which brand gets named

Here’s where my interest diverges from theirs, and the divergence is the whole piece.

Their dependent variable is the clickable citation, because that’s what a log can count. Mine is being named as the source, which arrives sometimes as a link, sometimes as your brand inside the sentence, and sometimes as both. Those aren’t the same measurement, so take their numbers as evidence about links and take what follows as an argument about names.

Read as an ecosystem story, ten-read-three-cited is a story about publishers going unpaid. Read as a competition story, it is a selection event with a survival rate of roughly a third, running on every query in your category, invisibly, all day. Six relevant sources were in the room and were not named. If you’re one of the six you have no idea, because from where you sit an answer that used your material and credited somebody else looks exactly like an answer that never touched you.

The authors open with Joan Robinson: the misery of being exploited by capitalists is nothing compared to the misery of not being exploited at all. For a brand, there’s a third state worse than either, which is being consumed while a competitor is named, because that answer is now actively selling against you using your own material.

Something chooses the three, and it is not a coin toss

The paper doesn’t ask what does the choosing. It isn’t their question and they say so.

Mine is a machine that has just read ten relevant passages and has room to credit three. It has to attach each citation to something nameable, and the ten pages are not equally nameable. Some belong to a source it holds as a resolved entity, with attributes, relationships and corroboration behind the name. Others are a URL with words on it. One of those gives the system an already-resolved candidate to name, and the other leaves more ambiguity about who or what should receive the credit.

For me, this is the reframe worth having: entity resolution has always been sold as a visibility problem, and it is really a selection problem, at the exact moment credit is handed out.

I know this from orchestrating models rather than from a study, and it is the same mechanism

I should be straight about the class of claim I’m making, because I’ve spent the last week insisting other people do the same.

This is reasoning from mechanism, not measurement. What I have alongside it is fourteen years of watching brands with resolved entities get named for things they barely wrote, and something more immediate: I orchestrate AI models all day, and the effect is not subtle. Give a model badly organised material and let it roam, and the work comes back poor. Organise the material properly and it comes back good. Organise it and tell the model where to look and it comes back exceptional, every time, without exception in my experience.

The retrieval and selection running inside an assistive engine presents the same underlying problem: choosing correctly among competing pieces of information, with nobody there to point. Organised, resolvable, plainly labelled material gets picked up and used correctly, and scattered material gets read and dropped.

Attribution has a supply side the engine controls and an allocation side you can influence

Split it into two problems and the strategy falls out, because the two belong to different people.

Supply is how much attribution an engine issues at all, and that is a product decision made inside the vendor: the 0.19 to 0.45 spread is what those decisions look like from outside, nothing you do moves it, and it will keep moving on its own as the products compete on trust and as regulators take an interest in exactly the telemetry Strauss and his co-authors are asking for. Allocation is which sources receive the attribution that does get issued, and several things feed it, including how far the system trusts the source, how fresh the material is, and in some cases whether there’s a licensing arrangement in place. Entity resolution isn’t the whole of allocation, it’s the part of allocation you own.

Brands get this backwards with impressive consistency. They treat supply as something to optimise, which is roughly optimising the weather, and treat entity resolution as a technical chore for somebody else, when it’s the foundational allocation input they can deliberately engineer.

Your server logs give you a rough proxy, and no dashboard is going to give you better

You can get closer to this than nothing, and a great deal further from it than you’d like.

Strauss and his co-authors could measure the gap because they had logs pairing what a system fetched against what it cited. You can run a crude version on your own infrastructure: count the assistive-engine crawlers hitting your pages, set that against the referrals and mentions coming back, and the shortfall tells you something real about how much you are being consumed relative to what returns. What it can’t tell you is the part that decides deals, which is which answers used your material and named a competitor, because the answer never reaches you and a crawler log says nothing about what happened next.

Every AI visibility product on the market, ours included, measures the half that surfaces: whether you were mentioned, whether you were cited, how you were described. None of them can show you the answers where you did the work and somebody else got the line, and any vendor claiming otherwise is selling you a number they cannot produce. Which is why the paper’s own recommendation is standardised telemetry rather than better commercial tooling.

Watch a brand decide to fight the ceiling instead: they publish more, more often, across more formats, which hands the engine more material to consume at their own expense while the thing that would have made them nameable stays undone. So the work is the unglamorous half, and it pays nothing on the day you do it: one version of every fact, everywhere it appears, matching, a home for the entity, third parties saying the same thing about you that you say about yourself. That is content work, so this was never a choice between publishing and entity building, and the difference between the two piles is consistency rather than volume. What it buys is being nameable, and nameable is what gets you into the three.

You can’t make the machine more generous. You can make sure that when it has three slots and ten candidates, you’re the one it can name without thinking. <!-WP:PRESERVED:1->

Publication note: the argument set out here is published for the first time on jasonbarnard.com on 29 August 2026. It builds on one measured component and one argued one. The measured component is “The attribution crisis in LLM search results: Estimating ecosystem exploitation” by Ilan Strauss, Jangho Yang, Tim O’Reilly, Sruly Rosenblat and Isobel Moure, Data & Policy, 28 April 2026, DOI 10.1017/dap.2026.10064, which defines the attribution gap as the difference between relevant pages an LLM reads and those it cites, and finds a ninety-two per cent no-citation rate for Gemini answers, approximately three relevant sites left uncited per query by Gemini and Perplexity Sonar, Sonar visiting roughly ten relevant pages while citing three to four, and a citation-efficiency range of 0.19 to 0.45 across models on identical queries, which the authors attribute to retrieval design rather than technical limits. That paper’s concern is economic: attribution, remuneration and the viability of content production. The claim published here is a different reading of the same finding. Attribution has a supply side and an allocation side: the engine decides how much attribution is issued, and a separate process decides which sources receive it. Where an engine reads ten relevant sources and credits three, that allocation is a competitive event with several inputs, including source trust, freshness and licensing, and the input available to the brand is whether the system has resolved it as an entity it can name. The paper’s dependent variable is the clickable citation; the claim made here concerns being named as the source, whether by link or by brand mention, and the two are not the same measurement. This claim is argued from mechanism rather than measured, and it is stated as such. A corollary is that a brand cannot audit its own position from outside, because an answer that used its material without naming it is indistinguishable from an answer that did not use it, which is why the paper’s own recommendation is standardised telemetry rather than better commercial tooling. Jason Barnard, Kalicube®.