What Google’s AI Contribution Pilot Says About Your Citation Reports
On September 14, Google confirmed it's testing a way to pay publishers when their content helps build an answer in AI Overviews, AI Mode, and Gemini (Search Engine Land). It's called the AI contribution pilot, it lives inside Search Console, and so far it's a small group of publishers. Digiday reported that at least dozens have been approached, and the early ones describe the payouts as "peanuts" next to their ad revenue (Search Engine Journal).
Most of the coverage has been about the money, which makes sense if you run a publication. I don't, and most of the people who read this blog don't either. You're on a brand team, and Google isn't going to pay you for anything. Still, the help text for the pilot has a line in it about what Google counts as "contributing" to an AI answer, and that line applies to anybody who reports on AI visibility, publisher or not.
What Google is testing
The short version, from the reporting so far. Publishers in the pilot see a panel in Search Console with a monthly earnings number and some history. Google hasn't said how that number is calculated, and one executive called it "quite black box." Publishers can opt out whenever they want. Google described it as an early-stage learning pilot, on top of the traffic and tools it already provides, which is corporate for "don't read too much into this yet" (Search Engine Land).
Then there's the part I care about. The payments only apply when a page shapes the answer while it's being generated. Google's help text says web content can also confirm facts or get linked inside its AI after the response has already been written, and "those instances don't qualify" (Search Engine Journal). So Google is drawing a line between two things most of us have been counting as one.
Two different ways to show up in an answer
Think about it this way. An engine gets a question, goes and reads some pages, and writes an answer from what it read. That's the first moment, and the pages it read are the ones actually shaping what the buyer hears. After that, it may go find sources that back up what it already said and attach them as links. That's the second moment, and a page that shows up there is decoration as far as the answer is concerned (useful decoration, maybe, since somebody might click it, but it didn't change a word the buyer read).
Two moments in an AI answer
An engine reads pages and writes the answer, then it may attach links that back up what it already said. Pick a view to see which moment each one looks at.
Based on Google's help text for the AI contribution pilot, as reported by Search Engine Journal, Sept 14, 2026. A page can appear in both steps; most reports can't tell which step a link came from.
A lot of AI visibility reporting is built on citations, which are the links. That's what's easy to see from the outside, so that's what gets counted. If you asked the people building those reports, they'd probably say, "Well, a citation is the best proxy we have," and that's fair, it is. Now Google has said in plain language that it doesn't treat a link and a contribution as the same thing, and Google can see both.
Citations move around a lot
The other problem with leading a report with citation counts is how unstable they are, even when nothing about your brand has changed.
Semrush and Kevin Indig ran the same 100 prompts across 20 buyer journeys through ChatGPT in two modes, minimal reasoning and high reasoning, and only 25.6% of the sources overlapped between the two modes, so roughly three out of four changed. The high-reasoning mode leaned a lot less on Reddit and reviews and a lot more on official documentation and government and academic sources (Search Engine Land).
I saw the same thing in my own data, where nothing at all changed between runs. In the AI Verdict Study I ran the same scripted buyer conversation six times per brand and engine, identical wording every time, and two identical runs shared only 38% of their sources on average (Where AI Engines Go to Decide About Your Brand).
Citations move around
Two views of the same problem. Change nothing about your brand and a big share of the sources still changes.
Sources: AI Verdict Study citation analysis (32,089 citations, 432 conversations, Sept 18 to 20, 2026); Semrush and Kevin Indig, 100 prompts across 20 buyer journeys in ChatGPT, as reported by Search Engine Land, July 1, 2026.
So if your citation count goes up 20% this month, I'd want to know a few things before anybody celebrates. Was it the same prompts? Same engine, same mode? Did you run it once or several times? In a lot of cases the honest answer to "why did it go up" is "we asked on a different day."
Where the shaping seems to happen
I should say up front that I can't see inside the generation step any more than you can. Nobody outside Google can, which is sort of the point of this whole post. You can still get a decent idea of which pages an engine is reading while it forms an opinion by looking at which sources show up at which point in a conversation.
When I sorted the 32,000 citations from my study by question, the brand's own website was 4% of the classified citations on the first question, the "what are my options" question where the shortlist gets built. Most of the reading at that stage was other people's pages: other companies, independent reviewers, and media. The brand's own site jumped to 35% and 37% once the buyer asked about tradeoffs and whether the product would still fit as they grew. Help center articles made up 29% of all the citations to a brand's own site, and 46% on the limitations question.
My guess is that those mid-conversation pages are much closer to what Google means by contribution than whatever gets attached as a link at the end. When a buyer asks what your product can't do, the engine goes and reads your help docs and repeats them, which is a very different thing from a roundup link sitting at the bottom of the response.
What I'd do about it
A few things, in order.
Number one, stop leading your AI visibility report with a citation count. Keep tracking it if you want, but lead with the things that describe the answer itself. Were you named? Were you recommended when the buyer asked? Were you described accurately? Run each check more than once too, since the sources can shift a lot from one run to the next.
Two, fix the pages engines read in the middle of a conversation, starting with your help center. I would urge you to find the articles that explain what your product doesn't do and rewrite them in plain language, knowing they'll get quoted to a prospect. Same for honest comparison pages. If you haven't written the comparison, the engine reads the one your competitor wrote.
Three, work on the pages that build the shortlist, which mostly aren't yours. Independent review sites, the trade media in your category, the round-ups. That's where the first answer comes from.
Four, if you do happen to be a publisher and you get the invite, read the fine print before you opt in. One publishing executive quoted in the coverage warned that taking the payments now could weaken publishers' hand in negotiating better terms later (Search Engine Journal).
Then last, keep an eye on that Search Console panel. Right now only a few publishers can see it, and all it shows is earnings, but it means Google has built something that measures when a page shapes an AI answer. If a version of that ever shows up for everyone, even as a simple count with no money attached, I'd want it at the top of any AI visibility report. I have no idea whether Google will do that, though I'd love it if they did.
Is your report measuring the answer or the links?
Five questions about how your team reports AI visibility. Answer all five to see where to start.
Nothing you click here is saved or sent anywhere.
One more bit of context, since some people will still ask whether any of this matters when the clicks are going away anyway. Ahrefs looked at 300,000 keywords and found that top-ranking pages now get 58% fewer clicks when an AI Overview shows up (Ahrefs, via Business Wire). If fewer people click, then what the answer says about you is doing more of the work, and it's worth knowing which pages are shaping it.
I wrote about a lot of this in Explainable, mostly the part about how engines decide what a brand is. The pilot is new, though, and I'm curious how other people are reading it. If you're reporting AI visibility to a CMO or a client, I'd like to hear what you're putting at the top of the report, especially if you think I've got this wrong.
Jarred Smith is the author of Explainable: Why AI Recommends Some Brands & Ignores Others, an Amazon bestseller on AEO, GEO, and SEO. He's a marketing leader with nearly 20 years of experience across healthcare, public media, retail, and environmental services. Find him at jarredsmith.com.