What actually decides who AI recommends

If a client asks an AI who to instruct and it names a firm, something on that list of sources was the reason. Most of what's cited alongside it wasn't. We read and judged thousands of those answers to find out which channel actually earns the pick, and the result is not close: comparison and ranking sites decide it roughly three times more often than the next best channel, Reddit, press or YouTube, and they are consistently the least worked on. That ordering held under two different tests, one generous and one built specifically to knock it down.

What actually gets you recommended

A best-providers article is pre-chewed for an engine, an ordered list with reasons already attached, so an AI asked who is best can lift the structure whole. That is why comparison and ranking sites come out on top whichever way the influence is measured, generous or strict. If you are deciding where to put effort first, this is the channel: highest return, least competition.

Two corrections to what we originally published, both worth knowing if you're prioritising channels. First, Reddit is not the clear second-place channel we said it was, under either test it sits level with press, not meaningfully above it. Second, this isn't a rate that holds evenly across engines. On Google AI Overviews specifically, YouTube is cited constantly and decides nothing.

Same on Gemini, ChatGPT and Brave, only Perplexity and Grok let video move the pick at all. A channel can be cited constantly on one engine and decide nothing there, while doing real work on another. That's the practical reason to never trust a blended cross-engine number: ours ranged from nothing to roughly half depending on the engine, and the average describes an engine that doesn't exist.

Who gets cited isn't who gets picked

Press dominates the citation count, cited in roughly twice as many answers as Reddit. If your visibility score is built on share of citations, press looks like the boss of the board. It isn't: press decides the pick far less often than it's cited, which is the gap between showing up and being the reason.

The question decides as much as the channel

Sort the data by the type of question being asked and a firm's presence in the answer is not one number, it is a spread. The chart shows how often the firm's own website appears among the cited sources, by question type, in the local professional-services market, the cleanest presence measure the data supports.

A fifty-fold swing driven purely by what the buyer typed. And underneath the thinnest bar sits the starkest verified number in the whole study: across 1,622 problem-led answers, “I have this specific situation, what do I do”, the question closest to a sale, the firm was not mentioned once. Not cited less, not ranked lower: absent, while the engine handed back generic explainers naming no provider at all. Measure your presence per question type, because a healthy average hides near-total absence from the questions worth the most.

Your own website is table stakes, not the decider

This is the one most worth checking, because it is where most budget goes. When a firm gets recommended, is its own website the reason? In the market where we labelled every answer, 817 of them: 106 where the firm was recommended, 694 where it was merely listed as one option, the firm's own site was among the cited sources about as often either way. Being recommended did not make its own site more present; statistically the two rates are the same number.

Your own site being cited does not separate the recommended from the also-ran, because it is there for both. The sharper version survived our re-check even more strongly than we first published: in 1,828 recommendation answers we could re-examine, the firm's own site was the only cited source not once. Zero. The sources that sat alongside it, the third-party layer, are where the difference between listed and recommended gets made. Owned content is the price of entry, not the thing that wins, and spending more on it whilst ignoring the third-party layer is spending on the part that was never in doubt.

The weight hides in a handful of URLs

Being named and being chosen are two different jobs, and a firm can be all over the answer and still rarely be the pick. In a 115-answer sweep of one UK advisory market, the leading firm converted 62% of its mentions into recommendations; a rival with half the mentions converted 20%. Where a single thread or a single publisher carries hundreds of citations, that is where influence concentrates. We can see the pattern in the data, but never who created any of it, so this describes a mechanism rather than accusing anyone.

What to do with this

• Work on comparison and ranking sites first. They come top under every definition we tested, at roughly three times the next channel, and almost nobody is working them.
• Never accept a blended cross-engine figure. The same channel ran from zero to about half depending on the engine, an average across them describes no engine that exists.
• Don't buy citation share as influence. Press is cited in twice as many answers as Reddit; that is presence, not persuasion. And a channel cited 1,216 times can still decide nothing.
• Measure presence per question type. A healthy average hid a firm's total absence from all 1,622 of the highest-intent problem-led questions.
• Treat your own website as table stakes. It is cited about as often whether you are recommended or merely listed, and in 1,828 re-checked recommendations it was never once the sole source. The work that changes the outcome sits off your own domain.
• Ask anyone selling you an influence number which definition they used and whether they can show you the judgments. Ours weren't recorded the first time. They are now, see below.

How rigorously we checked this

The first version of this page published a single influence ranking without keeping the per-answer judgments behind it, so nobody, including us, could check it. We re-ran the whole thing: 6,517 answers, every citation judged and every verdict stored with the model and rubric that produced it, then sent every “this decided the pick” verdict to a second, independent model whose only job was to refute it. It refuted 84–92% of them, not because the first pass was careless, but because “decided the pick” has a generous reading and a harsh one, and they land an order of magnitude apart. The ordering above (comparison sites first, Reddit not second, the per-engine zeros) is what survived both readings, which is why we trust it enough to publish.

• 6,517 answers judged, every verdict stored with the model and rubric that produced it.
• 8,783 “this decided the pick” verdicts sent to a stronger model told to refute them; 84–92% did not survive.
• 71% agreement between two independent judges on the same citations, the honest error bar on any influence figure, ours included.
• Reddit and YouTube were judged in full, not sampled; press and comparison sites from random samples of roughly 890 answers each.