Written by Jeremy Souffir Founder, JTS Tech Services
Here is the short version, and then the part that actually concerns you. On 14 September, Digiday reported that Google has begun paying a small, invited group of publishers when their content is used inside its AI answers — AI Overviews, AI Mode and Gemini. The feature is called AI Contribution, it appears as an earnings widget inside Search Console, and Google confirmed it as "an early-stage learning pilot to test how best to reward high-quality content". If you sell products or services rather than articles, you will almost certainly never be invited, the reported sums would not change your month if you were, and nothing about your site needs to change today because of it. We are writing about it anyway, because in order to pay for content Google had to define what counts as content doing something — and the definition it published draws a line straight through the middle of how nearly everyone has been measuring AI visibility all year.
The line is this. Your pages can earn when they contribute significantly to the answer while the answer is being generated. Content that confirms a fact after the fact, or that gets linked to inside Google's AI once the response already exists, does not qualify. Read that twice, because Google has just said out loud that both of those things happen, that they produce outcomes which look identical from where you are standing, and that it can tell them apart. A link under an AI answer is not evidence that your page was used to build it. Sometimes it is. Sometimes it is a footnote attached to something that was already finished.
What actually happened this week?
Taken from the reporting rather than from anyone's summary of it, and separated into what is confirmed and what is characterisation.
- The programme is called AI Contribution and it lives in Search Console. It was first spotted in April 2026; what changed this week is that publishers inside it started talking and Digiday published the details on 14 September.
- Participants get a widget in the Search Console dashboard showing a monthly earnings figure with history behind it. That is the whole interface, as reported.
- It covers content used in AI Overviews, AI Mode and Gemini. It is a licensing arrangement layered on top of ordinary Search, not a replacement for it.
- Entry is by invitation. At least dozens of publishers have been approached; no total has been confirmed, and per the reporting it reaches beyond news publishers. It is described as appealing more to small and mid-sized publishers than to large ones.
- There is no upfront fee, and a participant can cancel at any time from Search Console settings.
- Google's own statement is that this is an early-stage learning pilot to test how best to reward high-quality content, on top of the traffic and tools it already provides. That is the full extent of the official comment.
- Now the characterisation, which is not Google's and which we are repeating as other people's words: one publisher described the amounts as "peanuts" against advertising revenue, others called the numbers lowball, and an executive familiar with the programme called the calculation "quite black box". No payout formula has been published and we are not going to invent one.
Why would Google draw this line at all?
Worth a paragraph, because the reason tells you how durable the distinction is likely to be. The honest answer is that Google has not given one beyond rewarding high-quality content, and we are not going to put motives in its mouth. What we can say is structural: the moment you decide to pay for something, you are forced to define it, and vague definitions cost money. If Google paid for every link it placed under an AI answer, it would be paying for its own citation formatting. So it had to separate the content that changed what the answer said from the content that appeared beside the answer once it was written. That separation is not a policy choice about publishers. It is a consequence of the arithmetic, which is exactly why we would expect it to outlast this pilot whatever happens to the payments.
The wider context is not hard to see either, though the causation is inference and we will flag it as such. This is the same year in which the UK's competition regulator imposed publisher controls on Google's AI features, in which the EU reclassified ChatGPT as a search engine for disclosure purposes, and in which Google began reporting generative-AI impressions to site owners in Search Console. A pilot that pays for content sits comfortably in that sequence. Whether it was caused by it, nobody outside Google knows, and the reporting does not claim to.
Whose programme this is, and whether any of it is yours
Four limits, because otherwise someone will read the headline and ask a question in a meeting that this paragraph would have answered. First: this is a publisher licensing pilot. If your site is a store, a services business or a B2B company, you are not the intended participant, there is no application form, and the correct expectation is that nothing arrives. Second: the money is reportedly small and the calculation is undisclosed, so even for an invited publisher this is not yet a revenue line anybody should plan around — those are the participants' own words about their own cheques, not our estimate. Third: it is Google only, and only three surfaces. ChatGPT, Copilot and Perplexity have no equivalent, publish no such definition, and nothing here tells you how they assemble an answer. Fourth, and the one that matters most for what follows: the generation-versus-afterwards distinction is Google's internal accounting rule for a payments pilot. It is not a ranking factor, it is not a signal you can optimise a tag for, and there is no report anywhere that shows you which side of it your pages landed on. What it gives you is a confirmed fact about how these answers are built. That is worth a great deal and it is not the same thing as a lever.
Why does this matter if you will never see a cent?
Because of what almost every AI-visibility habit has been doing since January. The standard practice — ours included, in the early months — is to run prompts, look at the answer, and count whether your domain appears among the links. Tools sell that count. Dashboards trend it. Whole strategies are being set on whether the number goes up. What Google has just documented is that the number is a mixture of two different events, and it has been all along.
One of those events is that your page was part of how the assistant arrived at what it said. Your price, your delivery terms, your comparison table, your spec — something in it shaped the sentence the customer read. That is the outcome worth having, and the reason is obvious the moment you say it plainly: it is the only version in which the answer carries your facts rather than someone else's, whether or not anyone clicks. The other event is that a link to your page was appended to an answer that was already written from other material. That is not worthless. It is a link, it can be clicked, and it may well be how somebody finds you. But it did not influence the recommendation, and if the assistant recommended a competitor and cited you underneath, the citation is in some sense the record of your loss rather than a measure of your visibility.
On 1 September, writing about the new generative-AI impressions report in Search Console, we said that the report tells you that you appeared and not how — that being the answer and being a footnote under a competitor's recommendation both register as one impression, and that the difference between those two is the entire commercial question. That was a limitation we could describe but not resolve. Two weeks later, the same company has confirmed the distinction exists, is measurable from the inside, and is worth different amounts of money. We did not expect the confirmation this quickly and it changes the confidence with which we would now say it.
The two conclusions that both get this wrong
The first wrong conclusion is that citations no longer matter and the tracking should stop. That is an overcorrection and it throws away the only instrument most businesses have. A citation count is still the cheapest available proxy for presence, a fall in it is still worth investigating, and nothing published this week says otherwise. The correct adjustment is smaller and more annoying: stop treating the count as a score and start reading the answers themselves, because the sentence above the links is where you find out which kind of appearance you had. If the recommendation is a competitor and your link sits underneath it, your dashboard has been recording that as a win all year. The second wrong conclusion runs the other way, and it is the more expensive one: deciding that because Google is now paying for content that shapes answers, the move is to produce more content aimed at being ingested. That is 2011 content-farm logic wearing new clothes, and the pilot's own rule argues against it — the earning condition is contributing significantly to a specific response, which rewards pages that resolve a real question precisely, not pages that exist in volume. Both mistakes come from the same place: treating a newly visible distinction as a newly available tactic. Nothing became optimisable this week. Something became knowable, and the useful response to knowing is to measure differently, not to build differently.
What would we actually do this week?
- Pull ten real customer questions and read the whole answer, not the link list. Your best sellers, your highest-margin lines, the comparison people actually type. For each one write down a single word: recommended, mentioned, or cited. That three-way split is the measurement this week justified, it takes an afternoon, and no tool currently sells it to you.
- Go back through whatever AI-visibility reporting you already have and re-read its definition of a hit. If it counts domain appearances anywhere in the response, it is mixing the two categories. You do not need to throw it out. You need to write the caveat down beside the chart, once, where the person presenting it will see it.
- Check which of your pages actually answer a question in their opening lines. The earning condition is contributing significantly to a response, and the pages that do that are the ones stating a real price, a real delivery window, a real compatibility or a real limitation in plain text near the top. This has now been the answer under every study, protocol and report we have covered this year, which is starting to be less of a coincidence than a finding.
- Open the generative-AI performance report in Search Console and compare it against what you just read by hand. The report gives you volume with no quality dimension; your ten answers give you quality with no volume. Neither is worth much alone and together they are most of a picture.
- Do not go looking for AI Contribution in your Search Console. It is invitation-only, its absence tells you nothing about your site, and the time would be better spent on the first item in this list.
- Check your edge is not turning Google's crawlers away. We repeat this in nearly every post on this subject because it keeps being the actual cause of an empty report, and a page that cannot be fetched cannot contribute to anything, paid or not.
- Leave your budget alone. Nothing in this pilot is a channel, a ranking factor or a forecastable revenue line, and anyone proposing a spending change off the back of a licensing test they read about on Monday is proposing it off the back of a headline.
The genuinely encouraging part
Set the caveats down for a paragraph, because the underlying development is better than the cheque sizes make it sound. For two years the central complaint about AI answers has been that they are an unlit room: content goes in, an answer comes out, and nobody outside the building can say what was used or why. This week a company that runs one of those rooms published, in writing, a distinction between content that built the answer and content that was attached to it afterwards — and then did the one thing that makes a definition credible, which is attach money to it. You may never receive any of that money. You now know the distinction is real, which is strictly more than you knew on Friday, and knowing it costs you nothing. Better still, the work it points at is the work that was already worth doing. A page that states the price, names the limitation, answers the comparison in its first two sentences and loads for a crawler is the page that can contribute significantly to an answer. It is also the page a human in a hurry wanted. We have now written that sentence about a protocol, a regulation, a report and a licensing pilot in the space of a month, from four organisations with nothing in common. At some point the repetition stops being a coincidence and starts being the strategy.
Where we fit
The reason to bring us in over this is not the pilot, which affects you not at all and which you can safely ignore. It is the measurement gap it just exposed, and the fact that closing it is tedious rather than difficult. Knowing that recommended and cited are different outcomes is free. Finding out which one you are getting, across five assistants, for the products that actually pay your bills, means running the real questions, reading every answer rather than counting links, recording which competitor was recommended in your place and what was said about them, and then tracing each gap back to something specific — a page that never states a price, a comparison you have never written, a spec that exists only in an image, a bot rule at your own edge that turns the crawler away before any of it matters. That is what an AI Shopping Visibility audit is. We do that reading, set it beside your Search Console baseline so the two measurements check each other, and hand you a short ordered list of the things worth fixing first. Google has confirmed there are two kinds of appearance. We tell you which kind you have.
Sources
- Digiday — Google rolls out pay-per-value AI licensing program to publishers (the 14 September report and the primary source for the programme's existence and shape: the Search Console earnings widget with monthly and historical figures, the at-least-dozens of publishers approached, the appeal to small and mid-sized publishers over large ones, the opt-out, the "quite black box" characterisation of the calculation, the "peanuts" and lowball descriptions of the amounts — all of which are the participants' words rather than ours — and Google's confirmation that it is an early-stage learning pilot)
- Search Engine Journal — Google tests paying publishers for AI answers via Search Console (independent confirmation, and the source we relied on for the exact participant-facing wording of the rule this post is built around: that content qualifies when it "contributes significantly" during response generation, and that web content which "may also confirm facts or be linked to within Google's AI after a response is generated" does not qualify)
- Search Engine Roundtable — Google AI Contribution pilot — paying publishers for use of content in AI (the second independent write-up, the April 2026 first sighting of the feature, and corroboration of the same two lines of help text and of the cancel-any-time setting)
- Google Search Console Help — Generative AI performance report (Search) (the free report every site owner already has, referenced above as the volume half of the picture — impressions across AI Overviews and AI Mode, with no quality dimension and no click data)
- Google Search Console Help — Search generative AI control (the include/exclude setting that governs whether your content is eligible for these surfaces at all — the switch that has to be on before any of this is even a question)
- Google Search Help — What are impressions, position, and clicks? (Google's own definitions, including how links inside an AI Overview are counted — useful for seeing how flat the public metric is compared with the distinction the pilot draws internally)
- JTS Tech Services — Google now shows your AI impressions (1 September, and the post this one is a direct sequel to — where we said the impressions report cannot tell you whether you were the answer or a footnote, and that the difference is the entire commercial question)
- JTS Tech Services — What ChatGPT actually reads from your site (the equivalent mechanics on the other major assistant, and the reason the page-level advice above is about opening sentences rather than markup)
- JTS Tech Services — The EU just classified ChatGPT as a search engine (the disclosure track running alongside this one, and the other reason more of how these answers are assembled is likely to become visible over the next year)
