Generative Engine Optimization: What the Data Shows

Almost nobody publishes their numbers

Writing about optimising for AI search is mostly forecasting. The genre is full of confident advice about how ChatGPT and Perplexity choose sources, and very little of it is attached to data, because the people with the data are usually agencies who would rather sell the conclusion than show the working.

I have a dataset that is useful precisely because something went wrong.

I run a reference site, Role Players Guild, that grew to roughly 1,300 articles and 658 daily organic visitors in nine weeks. On August 22 it received a Google spam action, caused by my own publishing rate, and most of its Google traffic disappeared. That is a bad outcome and I have written about it elsewhere. It is also an unusually clean natural experiment: one search engine changed its mind about a site overnight, and every other engine kept reading the same pages, unchanged, on the same domain.

Here is what they did.

Fourteen days before, fourteen days after

Pageviews by referring domain, from PostHog. The left column is August 8 to 21, the right is August 22 to September 4. Same site, same content, same fourteen-day length.

Referrer	Aug 8-21	Aug 22-Sep 4	Change
Google	4,629	470	-90%
DuckDuckGo	1,917	3,257	+70%
Bing	940	1,418	+51%
Brave	176	369	+110%
Ecosia	142	268	+89%
ChatGPT	63	102	+62%
Qwant	49	96	+96%
Kagi	18	38	+111%
Direct	3,059	2,691	-12%

Every engine that is not Google grew. Several of them roughly doubled.

I want to be careful about what this does and does not show, because the obvious reading is too flattering. I did not do anything clever in those two weeks to earn that growth. Part of it is ordinary indexing progress on a site that was still young, and some of the increase is the second half of a trend that had already started. The point is not that a penalty causes growth elsewhere.

The point is the absence of a link. A judgement severe enough to remove 90 percent of one engine's traffic did not propagate to any other engine at all.

Your Google position is not your AI visibility

That non-propagation is the whole argument for treating this as a separate discipline, and it is worth being precise about why it happens.

These systems do not share an index. Bing runs its own crawler, and DuckDuckGo, Ecosia and Qwant all draw on Bing's results, which is why those four move together in the table above. Brave built a genuinely independent index. Kagi blends several sources. ChatGPT's retrieval does not depend on Google's ranking of a page.

So when Google demotes a domain, it is a statement about how Google's systems assess that domain. Nothing forces the others to agree, and in this case none of them did.

The practical version for anyone running a store or a content library: if your entire measurement of visibility is one Google Search Console property, you are watching a single engine's opinion and calling it the weather. On the site above, the fourteen days after the penalty still delivered more than 5,500 referral pageviews from non-Google engines.

The ChatGPT number is small, and that is fine

ChatGPT sent 102 pageviews in fourteen days. DuckDuckGo sent 3,257. Anyone telling you AI search has already replaced conventional search is not looking at referral logs.

Two things keep that number honest and both cut in the same direction.

Referral counts undercount AI influence structurally. A person who reads an answer, gets what they needed, and never clicks leaves no trace in your analytics at all. Some AI clients strip the referrer entirely, and those visits land in the $direct bucket alongside everything else. So 102 is a floor, not a measurement.

And even as a floor, it grew 62 percent in a fortnight while the site's dominant traffic source collapsed. Small and growing on an independent trajectory is a reasonable thing to prepare for. Small and growing is not a reason to rebuild your content strategy around it this quarter.

What actually makes content citable

Strip away the vocabulary and generative engine optimization is a retrieval problem. Something has to find your page, decide a passage in it answers a question, and quote it. Four things follow from that, and they are unglamorous.

Answer the question in the first paragraph

Retrieval systems quote passages. An article that spends 400 words establishing why the question is interesting before answering it has given an extraction system nothing to take.

State the answer plainly and immediately, then spend the article earning it. This is also better for humans, which is the useful thing about most of this work: the practices that make content extractable are mostly just the practices that make content clear.

Write sections that survive being taken out of context

Assume any section may be lifted with nothing around it. A section opening with "As mentioned above" or hanging on a pronoun that points three paragraphs back is unusable the moment it is extracted.

Concretely: descriptive headings rather than clever ones, self-contained first sentences, and units and time frames stated inside the sentence rather than inherited from a paragraph above.

Serve machine-readable text

Every page on this site has a plain markdown twin. Append .md to any article URL and you get the source, without the navigation, the JavaScript or the layout. There are 226 of them and they are listed in llms.txt.

The honest case for this is narrow. No crawler is obliged to read llms.txt, and I would not present it as a ranking factor. What it does is remove the parsing step between your content and anything trying to read it, at close to zero cost once it is wired up.

The unglamorous part is that infrastructure like this fails silently. When I sat down to write this article I checked that the twins actually worked in production and found that all 226 of them returned 404 on the live site while returning 200 locally. The file advertising them was fine, which is exactly what kept the fault invisible: the site was handing crawlers a list of URLs that did not resolve. The full write-up of that bug is worth reading if you plan to build the same thing, because the failure mode is more interesting than the feature.

Be the primary source for something

This is the one that is hard to fake and the one that matters most.

Retrieval systems reward the document a claim comes from. The table earlier in this article is not available anywhere else, because it is my own site's referral data during an incident I caused. That is a weak moat and it is still a real one. Summarising other people's numbers puts you one hop away from the citation; publishing your own puts you at the source.

You do not need dramatic data. Your own conversion rates, your own support ticket categories, your own before-and-after on a change you made: any of it is more citable than another restatement of a study everyone has already restated.

What does not appear to work

Named because they take up most of the oxygen in this topic.

• Keyword stuffing for AI. The systems doing the retrieving are language models. Density tricks aimed at term-frequency ranking are aimed at the wrong mechanism.
• Schema markup as a citation strategy. Structured data is genuinely useful, and this site ships plenty of it, but it earns rich results and entity clarity rather than LLM citations. Treat it as good practice, not as the lever.
• Publishing volume. I have the most direct possible evidence against this one. Rate is a constraint, not a target, and I spent a site's Google traffic learning it.
• Anything sold as a guaranteed placement in AI answers. Nobody controls that surface, and the vendors claiming otherwise are describing an ad product that does not exist.

How to measure it without fooling yourself

Two habits, both cheap.

Break referrals down by domain rather than trusting a bundled "AI traffic" segment. The table above only says anything because the engines are separated: bundled into one line, the most interesting finding in this article disappears. chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com are the domains worth watching, alongside the non-Google search engines.

And watch a set rather than a single number. One engine's referrals moving is noise. Eight engines moving in the same direction while a ninth moves the other way is a finding.

Where I would start

If you have one afternoon: take your single best-performing page, and rewrite its opening so the first paragraph answers the question it targets. Then open your analytics, break pageviews down by referring domain, and find out which engines are already sending you people. Most site owners are surprised, in one direction or the other.

If you have a week: add markdown twins and an llms.txt manifest, and check they resolve in production rather than only in development.

None of that requires believing any specific forecast about how search will look in 2028. It is defensible if AI search stays a rounding error, and it is a head start if it does not.

I do this work for e-commerce stores, mostly on Shopify. If organic traffic is something you want to stop renting, tell me what you sell.