Google now pays for AI citations. The fine print says a link isn’t a contribution.

On September 14, Digiday reported that Google has been quietly paying publishers through a new panel in Search Console called “AI contribution.” Opt in, and a monthly dollar figure appears when your pages help build answers in AI Overviews, AI Mode, and the Gemini app. Google’s statement calls it “an early-stage learning pilot to test how best to reward high-quality content, on top of the traffic and tools it already provides.”

The money isn’t the story. One source told Digiday the early returns were “peanuts” next to ad revenue. The story is the eligibility rule in the help text. Search Engine Journal’s read of it: a page earns only when it contributes significantly while the answer is being generated. Content that confirms facts, or gets linked after the response exists, doesn’t qualify.

Read that as a marketer, not a publisher. Google has just said, in a product with real money behind it, that citations come in two kinds. One shaped the answer. The other was attached to it. Only the first one counts.

Every visibility dashboard I’ve seen, ours included, reports them as the same number.

What Google actually built

The mechanics are thin, which is itself informative. Publishers get invited. They accept program terms in Search Console, a panel appears, and it shows earnings for the month plus some history. No query list, no page list, no formula. Barry Schwartz noted that it lives in Search Console rather than Publisher Center or AdSense. That’s the tool SEOs open every morning, not the one media executives use. The sign-up copy reads: “Accumulate earnings when your content contributes significantly to AI-generated responses in participating products.”

Digiday says at least dozens of publishers have been approached, that the pilot has proved more attractive to small and mid-sized outlets than to large ones, and that it has spread well past news. One executive with knowledge of the program summed up the reporting: “It’s quite black box.”

This isn’t Google’s first licensing move. Its June 18 policy post listed a News AI pilot with more than 200 publications, Extended News Previews payments to over 5,500 European titles, and News Showcase deals with 2,800-plus partners. Those are negotiated contracts with newsrooms. The AI contribution pilot is different in kind. It’s a metered payout, computed by Google, for any site whose content Google decides did the work. The same June post described the goal as partnering with “websites whose content meaningfully contributes to the freshness and factuality of generative AI responses” through grounding.

Freshness. Factuality. Grounding. Hold onto those three words, because they’re the closest thing to a published scoring rubric anyone outside Google has.

The line Google drew runs straight through your citation count

Google’s own developer documentation describes how AI answers get assembled. The AI features page says Search uses “query fan-out” to run related searches across subtopics, and then: “While responses are being generated, our advanced models identify more supporting web pages.” Supporting. Google’s word for the second wave of links is that they support the response, which is a different job from being the material the response was made from.

Researchers have a name for this split. A 2025 paper from Yash Saxena, Manas Gaur, and colleagues compares generation-time citation, where the model cites the sources it’s drawing on as it writes, with post-hoc citation, where an answer gets drafted first and sources are found afterward to back it up. They found the post-hoc approach gives broader coverage. It also means the link and the sentence never had to meet.

The pilot pays for the first kind. And you can’t tell them apart on the page. A grounding source and a support link render identically in an AI Overview: same chip, same spot in the carousel, same click. Your tracking tool sees a URL and counts a citation.

What you can observe	What it proves	Where it shows up
Impression	Your URL was eligible to appear	Search Console AI report (Google only)
Citation link	Your URL was attached to an answer	Every visibility dashboard
Contribution	Your content shaped what the answer says	Google’s payout ledger, for a few dozen publishers

We wrote in July that Search Console’s AI report shows impressions only, and that an impression is eligibility, nothing more. This pilot sharpens that argument. Google is now computing the bottom row of that table for real, and paying on it. It shows the result to a handful of sites as a dollar figure with no breakdown, and to everyone else not at all.

Why this matters even if you’ll never see a check

Three reasons, in order of how much they should change your Monday.

First, some of your citations are worth nothing, and you don’t know which. We’ve covered the fact that 61.7% of AI citations never say the brand’s name. The survey of 45 GEO studies we summarized this month put fully supported sentences in cited answers at 51.5%, and citations that correctly back the sentence they’re pinned to at 74.5%. Those numbers already said a link and an influence are different things. Google has now built a payment system on that exact distinction. When the company generating the answers decides “linked afterward” is worth $0 and “shaped the generation” is worth money, the fidelity argument stops being academic.

Second, the number you’d most want exists and is being withheld. Publishers in the pilot see a monthly total. They don’t see which pages, which prompts, or what “significantly” means. Everyone outside the pilot sees nothing. So Google now has two AI reporting surfaces, and both show an output with the inputs stripped out. I don’t think that’s an accident. Query-level contribution data would let sites reverse-engineer what grounding rewards. A dollar figure can’t be reverse-engineered by anyone.

Third, and this is the one for brand teams, the publishers who write about your category just got a paycheck tied to grounding. The paycheck isn’t tied to traffic or clicks. It’s tied to whether their page was the raw material for an answer. Google’s stated criteria are freshness and factuality. So the review sites, trade publications, and comparison blogs that already produce most of the citations AI engines use now have a direct financial reason to keep product facts current and specific. That’s the pool AI answers about your brand get built from. Expect it to get more factual and more frequently updated.

What “contributes significantly” probably rewards

Google won’t publish the formula. But someone ran the experiment at scale, and it’s a decent proxy for what generation-time citation responds to.

At SIGIR 2026 in July, Rahul Vishwakarma, Shushant Kumar, and Ratnesh Jamidar presented “What Gets Cited: Competitive GEO in AI Answer Engines”. They built 252,000 paired trials across six models, including Gemini 2.5 Flash, GPT-5.2, and Claude 3.5 Sonnet. Each trial handed the model two nearly identical product pages that differed in exactly one of 18 factors, with brand names swapped for fictional ones and source order counterbalanced, then recorded which page got cited first.

Four factors were gatekeepers on every model tested. The odds ratios are so large they stop being useful as numbers and become a verdict.

Factor	Effect on being cited first	Tier
Topic match vs. mismatch	OR above 10,000 on all six models	Gatekeeper
Price stated vs. not stated	OR above 10,000 on Gemini and Claude; 6 to 36 on GPT models	Gatekeeper
Recent vs. old timestamp	OR above 10,000 on Gemini, Claude, GPT-5.2; 69 to 1,494 elsewhere	Gatekeeper
Higher vs. lower list position	OR above 10,000 on Gemini, Claude, GPT-5.2; about 1,800 to 2,000 elsewhere	Gatekeeper
Claims with evidence vs. without	OR from 2 to above 10,000	Secondary
Direct vs. hedged language	OR from 2.7 to 754	Secondary
Content structure and layout	OR 0.79 to 1.68, no effect	Negligible

Look at Google’s two words again. Freshness is the third row. Factuality is the second and fifth: a page with a price on it and claims with evidence behind them is a page a model can ground on. The formatting tricks that fill GEO checklists came out flat. The authors’ own advice is to fix the gatekeepers first (topic, price, dates, position), then add specs and comparisons, and push formatting to the bottom of the list.

The usual caveat applies. The testbed put two documents in front of the model. Production systems retrieve five to ten or more, and the authors say so. But if you want a working theory of what earns a contribution payout versus a support link, this is the best evidence on the table, and it agrees with what Google said in June.

What to do this quarter

Split your citation reporting into two buckets. Pull a sample of answers that cite you across engines. For each one, ask whether the sentence next to your link contains something only your page says: a number, a date, a spec, a named claim. If a competitor’s URL could sit in that slot without changing a word of the answer, log it as a support link, not a contribution. Report the ratio. It’ll be uglier than your citation count, and more useful.
Put the paid-for facts on the page, high up. Explicit prices, dated updates, specifications, and claims with a source are what the SIGIR study found models ground on and what Google says it pays for. A page that’s on topic but hedged and undated is a page built to be linked after the fact. This is the same lesson as stale pages losing citations, with a price tag now attached.
Feed the publishers who are now paid to ground. If a trade site or reviewer earns when their page shapes a Gemini answer, they need fresh, specific, verifiable material from you. Wire releases mostly don’t get retrieved; the reporter’s page does. Give reporters the spec sheet, the price change, the dated benchmark, and the quote with a name on it.
If you run a content property, watch your Search Console settings. The pilot has spread beyond news, and the invitation shows up there. If it arrives, treat the monthly figure as a contribution signal you can correlate with your own citation tracking, not as revenue. And weigh the warning one publishing executive gave Search Engine Journal: accepting a metered payout may weaken your hand in any later negotiation.
Stop waiting for Google to show you the number. It has computed contribution, priced it, and declined to expose it. The only version of this metric you’ll get is the one you build: answer text, not link lists, sampled repeatedly, across every engine, Google’s included.

The line “citations are the new backlinks” had a good run. Google just priced it. A backlink was a vote. A citation, it turns out, is sometimes a source and sometimes a footnote added later, and the one company that can tell the difference is paying for the first and ignoring the second. Your dashboard should learn to tell them apart too.

Want to know what AI platforms say about your brand? Try RivalHound free and find out.