The short version: most of what is sold as GEO, AEO, or "AI SEO" has no independent evidence behind it, and some of it is explicitly contradicted by the search engines themselves. Below is what the research actually says, with every claim linked, including the parts that are inconvenient for anyone selling these services. We have a stake in this too, and we will tell you where we got it wrong on our own site.
The test we apply before doing anything
Before we do any work meant to earn AI citations, we ask one question: is there independent, ideally causal, evidence that this moves the needle? Most popular tactics fail that test. A few pass. The difference matters, because the failing ones are the easy, billable, technical fixes, and the passing ones are slow and mostly happen somewhere other than your website.
Start with what Google says, because it is the strongest source available
Google's own Search Central documentation on AI features states it plainly:
"There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary."
The same page adds that you do not need to create machine-readable files or AI text files, and that there is no special structured data to add. In May 2026 that guidance was revised to group llms.txt, content chunking, AI-specific rewriting, and special schema together as tactics that do not help.
When the platform you are optimizing for publishes documentation saying the optimization is unnecessary, that is worth more than any vendor case study.
What the evidence says does not work
Schema markup, as a citation lever. Ahrefs ran an actual causal test: 1,885 pages that added JSON-LD, matched against controls drawn from six million URLs, difference-in-differences, over 30-day windows. AI Overview citations went down 4.6%. AI Mode and ChatGPT moved 2.4% and 2.2%, both indistinguishable from noise. Cited pages do have more schema, but that is correlation. Keep schema for Google rich results and basic hygiene. It is not a way to get cited.
llms.txt. Ahrefs checked 137,210 domains. About 28% publish a valid file. 97% of those files received zero requests in a full month. No major AI lab has committed to reading it. My favorite detail in the entire study: of the fetches that did arrive, 12% came from GEO and AEO tools auditing each other. The industry is largely studying itself. We publish one because it is free and harmless, not because it does anything.
Keyword stuffing. The Princeton GEO study measured it at roughly negative 10%. Writing for the model the way people wrote for 2012-era Google actively hurts.
"AI visibility rankings." SparkToro ran 2,961 prompts across ChatGPT, Claude and Google with 600 volunteers and found less than a 1 in 100 chance of getting the same brand list twice, and roughly 1 in 1,000 of the same order. Rand Fishkin's conclusion was blunt: any tool claiming to give you a ranking position in AI is not measuring something stable. If a dashboard shows your "AI rank" moving week to week, a large part of what you are seeing is variance.
What the evidence supports
Being readable without JavaScript. Vercel and MERJ analyzed 569 million GPTBot fetches and 370 million from Claude: none of the major AI crawlers execute JavaScript. They download it and do not run it. They also found ChatGPT hitting 404s on nearly 35% of its fetches, against about 8% for Googlebot. So: your claims need to exist in the raw HTML, and your status codes and canonicals need to be correct. This is engineering, not marketing, and it is the part most SEO work never touches.
This one bit us. Our own homepage statistics used to animate upward from zero with JavaScript. A crawler reading the raw HTML saw the starting value. Our headline stat, the one about how many AI pilots return nothing, was being served to ChatGPT and Claude as "0%". We were publishing the exact opposite of our own argument to the exact machines we wanted reading it. We found it during an audit of our own site and fixed it. The number in the HTML is now the real number, before any script runs.
Statistics, quotations, and citing sources. The Princeton GEO work (KDD 2024, 10,000 queries) found real gains from adding statistics and citing authorities, with the largest effects for lower-ranked content. But read the caveat, because it matters: a critical survey of 45 GEO studies published in July 2026 found those gains are "conditional on a source already being present in a fixed context," and concluded that no reviewed technique shows a stable, cross-platform causal effect on organic discoverability. In other words, the "+40%" figure that anchors most GEO sales decks does not generalize the way it is sold. It describes what happens once you are already in the running.
Off-site presence. This is where the evidence is strongest and where almost nobody wants to sell you work, because it is slow. Brand mentions across the wider web correlate far more strongly with AI citation than backlinks do, and the large majority of citations trace back to third-party sources rather than your own domain. If your entire strategy lives on your own website, you are playing a small part of the game.
Bing, not just Google. ChatGPT's search leans on Bing's index, and most B2B sites never set up Bing Webmaster Tools or IndexNow. That is a fixable gap that has nothing to do with content quality.
The part nobody selling this will tell you
AI referral traffic is currently very small, and the best independent measurement says it does not convert better.
A peer-reviewed study in Marketing Science (April 2026) looked at 973 ecommerce sites with $20 billion in combined revenue and twelve months of first-party data: more than 50,000 ChatGPT transactions against 164 million from traditional channels. ChatGPT referrals came to roughly 0.2% of sessions, and organic search converted about 13% better. The authors note the results "contradict widespread expectations of LLM superiority."
Meanwhile marketers estimate AI drives about a quarter of their traffic. Measured, it is around 1%. That is a 24-fold overestimate, and it is the gap most GEO pitches are quietly selling into.
You will see claims that AI traffic converts 4x or 20x better. Those come from vendors, usually measured on their own unrepresentative niches. When one agency published a null result against its own commercial interest, organic and LLM traffic converted within a rounding error of each other, p = 0.794.
What a real first-party measurement looks like
Everything above is other people's research. Here is a measurement of our own, with the caveat that matters most placed before the numbers rather than after them: it is not from this site.
purviewx.ai is weeks old, and we have not yet run a citation probe against it. So we have no data about our own domain, and we are not going to borrow a number measured somewhere else and let it read as ours. That is worth stating plainly, because an earlier version of this section did exactly that, and it is precisely the move the rest of this post exists to criticise. We caught it in an audit of our own content and corrected it here.
What we do have is a consumer information site we built and operate for a client, which has been live long enough to measure. We probe it with a fixed set of questions a real user might ask, run repeatedly against multiple assistants, logging whether the site is cited. Across all probe questions, Claude cited it in 10 of 21 runs (48%), and OpenAI in 2 of 21 (about 10%). Separately, in a 24-hour window with search crawlers excluded, AI crawlers accounted for roughly 8% of the reads of its content pages.
Two honest caveats, and they matter more than the numbers.
We are not going to quote you the best result. On that site's single strongest question, it was cited in seven runs out of seven. That is a real measurement and it would look excellent on a slide. It is also one cell out of six, and those seven runs are the same prompt repeated, which measures stability, not coverage. Quoting it as a "100% citation rate" would be exactly the kind of thing this post is arguing against. The honest figure is 48% on 21 runs, and the 95% confidence interval on that is 28% to 68% — wide enough that the difference between a good quarter and a bad one would disappear inside it.
The crawler number is only meaningful because of what we took out of it. Bingbot is deliberately excluded from the AI bucket. Leave it in and the AI share inflates substantially, which is a mistake we made in an earlier version of this measurement and had to correct. If a vendor quotes you an "AI crawler share" without telling you what they classified as an AI crawler, the number means nothing.
That is the whole point. These are small numbers on small samples, and we are giving you the interval, the exclusion rule, and the name of the thing being measured rather than a headline. Anyone selling you a clean, confident AI-visibility percentage is either measuring something we do not have access to, or not telling you what we just told you.
So what should you actually do
Fix the things that are true regardless of how AI search evolves. Make sure your pages render their content in HTML without JavaScript. Fix your status codes. Get into Bing's index. Publish things worth citing, with real numbers and real sources. Build a presence somewhere other than your own domain. Then measure honestly, and accept that a single run of a prompt is variance, not a ranking.
Do not buy a monthly retainer whose deliverable is a visibility score. That number is available from a $99 tool, and the research above suggests it is not stable enough to manage against.
Why we are publishing this
None of it is proprietary. It is independent research anyone can read, and we have linked all of it. The reason it rarely leads a pitch, ours included until recently, is that it is inconvenient: the highest-leverage work is off your own site, slow, and hard to package as a technical deliverable. Selling a schema audit and an llms.txt file is easier.
We would rather tell you what the evidence says, including the part where we served our own best statistic to AI crawlers as a zero. That is the same discipline we bring to everything else: find out whether it actually works, including our own tactics, and drop the ones that don't.
PurviewX is embedded AI leadership for companies sitting on real operational data. We find out whether your AI actually works, including ours. Start a conversation.