Here is a number that should reframe how an affiliate publisher thinks about its job. When eMarketer looked at how ChatGPT discussed the eyewear brand Zenni, it found that almost 70 percent of the sites the model cited were affiliate marketing content. Not the brand's own site. Not a press release. Reviews, comparisons, and buying guides written by publishers who earn a commission.
That figure tells you two things at once. The first is reassuring: AI assistants are not skipping affiliate content. They are built on it. The reviews and roundups publishers have produced for twenty years are exactly what a model reaches for when a shopper asks it to compare options. The second is uncomfortable. Being read by the model and being recommended by it are different outcomes, and a shopper who gets a synthesized answer never clicks the link the answer was assembled from.
Generative engine optimization, usually shortened to GEO, is the practice of making content the model chooses to cite. For affiliate publishers it is not a side project. It is the new version of the core skill. This piece is the tactical how-to: what makes a review extractable, how GEO differs from the SEO publishers already know, and how an affiliate content operation should change to compete for the citation rather than the click.
Origin: where GEO came from
The term has a clean starting point. In late 2023, researchers from Princeton, Georgia Tech, the Allen Institute for AI, and IIT Delhi published a paper titled "GEO: Generative Engine Optimization." It went on to KDD 2024, a serious venue in data science. The paper did what the marketing blogs had not: it tested, at scale, what actually makes a generative engine cite a source.
The team built a benchmark called GEO-bench, 10,000 queries spread across domains, and ran a set of content tweaks to see which lifted a source's visibility inside an AI answer. The results were specific. Adding direct quotations from credible sources raised visibility by roughly 39 percent over the baseline. Adding relevant statistics lifted it about 31 percent. Citing sources lifted it about 29 percent. Improving the plain fluency of the writing added around 23 percent. At the bottom of the list, keyword stuffing, the oldest SEO trick, did nothing useful and edged slightly below the baseline.
That last detail is the real lesson. GEO was born as a correction. For two decades, search optimization rewarded moves that had drifted away from quality: keyword density, thin pages built around a search term, link schemes. The Princeton work showed that generative engines reward almost the opposite. They reward content that reads like it was written by someone who knows the subject, backs claims with numbers, and points to where the numbers came from. The honest version of content marketing turned out to be the optimized version.
Present: what makes affiliate content extractable
An AI assistant does not read a review the way a person does. It retrieves passages. When a shopper asks for the best budget mattress, the model runs searches, pulls a capped number of chunks from each candidate page, and assembles an answer from the chunks that best match the question. Your 4,000-word review is not evaluated as a whole. It is mined for the few passages that answer the prompt cleanly. GEO for affiliates is mostly the work of making those passages easy to find and safe to quote.
Several moves do the heavy lifting.
Lead with the verdict. Generative engines apply strict retrieval limits per page and favor information placed near the top. A review that buries its recommendation under 800 words of preamble is handing the model nothing to grab early. State the conclusion plainly in the first paragraph: which product wins, for whom, and why. Then support it. This inverts the classic affiliate structure that saved the verdict for the end to keep readers scrolling past the ads. The model does not scroll.
Make passages self-contained. A chunk the model retrieves has to make sense on its own, without the paragraph before it. A sentence like "it also runs cooler" is useless out of context. "The Cocoon Chill ran four degrees cooler than the Nectar in our overnight thermal test" survives extraction because every fact it needs is inside it. Write the key claims so each one could be lifted into an answer and still be true and clear.
Build real comparison tables. A table with named columns is one of the most extractable formats there is. Price, key specs, who each option suits, a one-line verdict per row. A citation-pattern study published by Search Engine Land found that listicles captured 40 percent of citations on commercial-intent queries, nearly double any other content type, because they break a decision into discrete, comparable units. A table is the most machine-readable way to express a comparison, and comparison is the exact job the shopper handed the assistant.
Put original data in the writing. This is the move that AI cannot fake, and it is where affiliate publishers have a structural advantage over the model itself. An assistant can summarize ten reviews. It cannot run the eleventh test. Numbers from first-hand testing, battery drain measured over a week, decibel readings, the actual return process timed and described, give the model something specific and verifiable to quote. Cyrus Shepard's review of 54 AI-citation studies and experiments found that factually specific, quantified statements consistently outperform vague claims, and a Digital Applied analysis of 1,000 AI Overviews found that pages carrying at least one named-source citation in the body were cited about 2.1 times as often as pages with none. Original testing data is the highest form of that signal, because the publisher is the source.
Answer the sub-questions explicitly. Models fan out. A query about the best running shoes spawns quieter searches about durability, sizing, and price. A review with clear headings that name those sub-questions, "How long do they last?", "Do they run small?", gets pulled into more of those fan-out answers. Write headings as the questions a buyer actually asks, then answer each in the first sentence under it.
Keep it current, and say so. Freshness is a strong citation signal, and AI engines lean harder on it than classic search does. Ahrefs, analyzing 17 million citations, found AI-cited pages were on average about 26 percent fresher than standard organic results, and that ChatGPT in particular favored URLs hundreds of days newer than the pages Google ranked. Other analysis points to a citation shelf life measured in weeks, not years, for fast-moving categories. For an affiliate publisher this means a visible, accurate "last updated" date and genuine updates behind it: refreshed prices, retested winners, discontinued products removed. A two-year-old roundup with stale prices is not just unhelpful to a reader. It is invisible to the model.
What does not work, and the schema question
GEO advice has its own myths, and one is worth correcting directly because publishers are spending money on it.
Schema markup, the structured JSON-LD code that has helped pages win rich results in Google for years, is widely sold as a GEO requirement. The evidence is thinner than the sales pitch. Ahrefs published a controlled study in May 2026, identifying nearly 1,900 pages that added JSON-LD schema and matching them against 4,000 control pages. The citation change in ChatGPT was about 2.2 percent, statistically indistinguishable from zero. In Google AI Overviews the change was actually negative, a drop the study found unlikely to be chance. Adding schema did not buy a single page more citations. Large language models do not parse schema the way a search crawler does; they read the rendered text. Shepard's analysis reached a similar verdict, scoring structured data as a small and inconsistent factor.
This does not mean schema is worthless. It still earns rich results in classic search, still feeds knowledge panels, still helps machines disambiguate entities. Keep it for those reasons. Just do not treat it as the lever that gets you cited. The lever is the text. A publisher choosing between an afternoon on schema and an afternoon testing a product should test the product every time.
The same realism applies to AI-generated review content. It is cheap to produce a thousand AI-spun comparisons, and that is the problem. When every publisher can generate the same synthesized roundup, none of it is distinctive, and the model has no reason to prefer one over another. The content that survives exists only because someone did the work: bought the product, used it, measured it, wrote down what happened. AI lowered the cost of average affiliate content to near zero, which means average affiliate content is now worth near zero. Original testing is the moat.
How GEO differs from the SEO publishers already know
GEO is not a wholly new discipline. The overlap with good SEO is large. Crawlability, topical authority, a fast site, content that matches search intent, all of it still matters, because assistants run real searches to ground their answers and an Ahrefs study found that pages cited in AI Overviews still skew toward ones that rank. The foundation is shared.
But four differences change the daily work.
The unit of success moves from the click to the citation. SEO optimized for a ranking that produced a visit. GEO optimizes for a mention inside an answer that may produce no visit at all. That forces a publisher to measure differently: tracking how often its brand and its reviews are named across ChatGPT, Perplexity, Gemini, and AI Overviews for the prompts buyers actually use, rather than watching only rank and sessions.
The unit of content moves from the page to the passage. SEO thought in pages competing for a keyword. GEO thinks in extractable chunks competing to be quoted. A page can be cited for one excellent table while the rest of it is ignored. That rewards reviews built as a series of self-contained, well-labeled answers rather than one long flowing argument.
Format follows intent more strictly. Ahrefs found that for commercial queries, AI answers cite listicles and structured comparisons disproportionately, and for informational queries they lean on articles. A "best X for Y" review needs to actually be structured as a ranked comparison, not an essay that happens to mention several products.
And the competitive set widens. In classic search a publisher competed with other publishers for ten slots. In an AI answer it competes with everything the model read, including Reddit threads, YouTube transcripts, and the brand's own site, for a handful of citations. Ahrefs found that only 38 percent of AI Overview citations now come from pages ranking in Google's top ten, down sharply from 76 percent months earlier. Ranking first is no longer close to sufficient. The answer is assembled from a much larger and messier pool.
Future and impact: rebuilding the affiliate content operation
If the citation is the prize, an affiliate operation built for the click needs to change in concrete ways.
Reorganize the editorial template. Every review should open with a clear verdict, carry a real comparison table, use headings written as buyer questions, and surface original testing data in the body. This is a template change, not a tone change, and it can be applied to a back catalogue as well as new posts.
Build a testing function, not just a writing function. The defensible asset is first-hand data, so the operation needs a way to acquire and test products and record measurements. For a small publisher that might be one disciplined process. For a larger one it is a lab function. Either way, the budget shifts from volume of content toward depth of evidence per piece.
Treat updating as a permanent job. Because freshness drives citation and prices move constantly, a fixed schedule for refreshing prices, retesting winners, and pruning dead products becomes core operations, not an afterthought. The "last updated" date has to be true.
Measure citation share. The dashboard a publisher watched, rank and organic sessions, no longer captures the value. The new measurement is presence: across a fixed set of category prompts, how often the model names your brand and pulls from your reviews, and how that compares to competitors. A small set of tools now does this monitoring at scale, the closest thing to a rank tracker for the answer era. Affiliate publishers are among the most exposed to this shift, so the measurement cannot wait.
There is a harder question underneath all of this. Getting cited is not the same as getting paid. A model that quotes your review and recommends the product, but sends the shopper straight to a checkout with no tracked link, has used your work without compensating it. That is a real and unresolved risk, and whether a tracked affiliate link can survive an agent-mediated purchase is its own open debate. GEO does not solve that problem. What it does is keep a publisher in the conversation. A review the model never reads earns nothing for certain. A review the model relies on at least has standing to argue for credit, and standing is the precondition for everything else.
The publishers who treat this as a sales motion will lose. The ones who treat it as a return to genuine, tested, well-structured product journalism are building exactly what the models reward, and what a human reader wanted all along. The skill was always supposed to be helping someone choose well. GEO just removed the option of faking it.
Council summary
This post argues that affiliate publishers should optimize for being cited inside AI answers, not just ranked, and that the winning moves are the honest ones: a verdict up top, self-contained passages, real comparison tables, original testing data, and genuine freshness. The accuracy pass verified the headline figures against their sources: the roughly 70 percent affiliate share in ChatGPT's Zenni answers (eMarketer), the Princeton GEO-bench results across 10,000 queries, the Ahrefs freshness and top-ten citation studies, and the May 2026 Ahrefs schema test. Three corrections were made: the statistics-addition lift was set to 31 percent to match the Princeton table, the 2.1 times named-source claim was reattributed from Ahrefs to the Digital Applied analysis of 1,000 AI Overviews that actually produced it, and an unsupported Impact.com claim about structured reviews was replaced with the Search Engine Land listicle finding. The schema passage was also sharpened so it no longer understates a result the study found significant. The takeaway for a busy publisher is concrete: rebuild the editorial template around extractable, tested, current reviews, and track citation share, because faked content no longer competes.
Comments