The evidence, including
what did not survive.
Most of what this category sells has never been shown to work. We went and checked before we built a service on it, and the checking cost us several things we would have liked to tell you.
The number that shapes the offer
What happens to an AEO result when you add a control group?
It mostly disappears. When 1,885 pages added schema markup, citations rose 43% — until the same window was measured on matched pages that changed nothing, which rose almost as much. The honest figure is +2.4%.
Raw before and after, treated pages only
+43%
Same pages, after matched control pages
+2.4%
The refusal list
What will we not sell you?
Six things you will be quoted for elsewhere. Each row is what we found when we looked for the evidence, with the source next to it.
| Sold as a lever | What we found | Source |
|---|---|---|
| Schema markup, to get you cited | No lift. After matched controls the effect was +2.4% on AI Mode, +2.2% on ChatGPT, and −4.6% on AI Overviews. | Ahrefs, 11 May 2026 |
| An llms.txt file, so models can read you | Google states plainly that Search ignores these files, and that keeping one neither harms nor helps. | Google, AI features and your website |
| AI-specific markup, Markdown mirrors, chunk-sized rewrites | Google states you do not need to create new machine-readable files, AI text files, markup or Markdown to appear in Search. | Google, AI features and your website |
| A guaranteed citation, or a top-five placement in the answer | Nobody outside Google can see the mechanism being guaranteed. Google says so in its own documentation. | Google, AI features and your website |
| A number for how often your buyers ask ChatGPT a given question | No engine publishes prompt frequency. Every figure on the market is reconstructed from an opt-in panel or extrapolated from Google volume. | Our own check, September 2026 — no published source exists to link |
| "AI traffic converts 42% better" | We could not retrieve the underlying report. Every version of the figure we found was relayed secondhand, and the comparison baseline was all non-AI traffic, not organic search. | Our own check, September 2026 |
What did hold
Is anything about this actually established?
Three things, and they are enough to justify the work. Two of them come from the platform itself, and the first of them is why what actually changes between SEO and AEO is a shorter list than the category implies.
“Optimizing for generative AI search is optimizing for the search experience, and thus still SEO.”
“No third-party tool has access to our internal ranking or AI systems.”
Users clicked a traditional result on 8% of searches showing an AI summary, against 15% without one.
What that leaves
So what can anyone actually promise?
Three layers, and the promise gets weaker as the layer gets further from what we touch. Every guarantee this practice makes attaches to a specific one of them.
01Guaranteed
we control this
Access and extractability
Whether an engine can fetch you, render you, and lift a passage out of you. Our work decides it; a re-test proves it.
- Crawler and edge policy, fetched from outside your network
- Complete in raw HTML, with real links and stable anchors
- One entity string everywhere, resolved in schema
The audit invoice is cancelled if it surfaces no cause here you could not reasonably have found yourself.
02Committed
we ship into it
Answer coverage and evidence
Whether the passage an engine needs exists at all. We control what ships. We do not control what gets chosen.
- Question-headed passages that survive being pulled out
- Comparison and fit pages for the questions buyers ask
- Independent corroboration where models actually read
A count of passages and corrected sources, agreed before kickoff and reported against monthly.
03Measured
we only report it
Whether the engine names you
Engine behaviour. Nobody controls it, and anyone quoting a citation guarantee is selling a number they cannot reach.
- Named rate on a panel frozen before kickoff
- Who was named instead, and on which sources
- Every factual error an engine stated about you
Reported, never promised. What we guarantee is the traffic it brings: for qualified brands, no management fee until US organic traffic is up 30%.
Before you pick anyone
What should you ask an AEO vendor?
Eight questions, and the answer that should satisfy you. Ours are in the third column, including the two we fail.
| Ask any vendor | A good answer | Ours |
|---|---|---|
| Who else have you done this for, and can I call them? | A brand name and a person who will take the call. | No AEO client references yet. Ask this first, of us and of everyone else — it is the strongest evidence any agency can give you. |
| What is your own AI visibility? | A number they publish, whichever way it reads. | 13–19 October 2026: named in 528 of 1,050 AI answers, 50.3%, first of every entity — up from 4.7% in the first week on the same tool. Published monthly on our results page, where every change of instrument is reported as the break it is. |
| What tool produces the number my fee depends on? | Named, with the raw responses handed over. | Our own runner, model and version recorded per observation, raw responses archived and handed to you at the end. |
| Will you fix the question set before you start? | Yes, and you get the list. | The set is chosen with you from a week of live readings, frozen before any work starts, and unchanged for twelve months. Adding questions later moves the number without moving the business. |
| What is the baseline, and on what date? | A figure and a calendar date, before any work. | Month 0 is a measurement month with no changes shipped. Every later report is read against it. |
| What exactly do you guarantee, and what happens if it fails? | Something the vendor controls, written into the agreement. | The audit invoice is cancelled if it surfaces no cause you could not reasonably have found yourself. For qualified brands, no management fee until US organic traffic is up 30% on a season-adjusted baseline, measured at day 90. |
| Who verified your published result? | The client, on the record. | No published result yet. The first one will carry the client’s attestation or it will not go up. |
| Can you guarantee a citation? | No. Anyone who says yes is selling a mechanism they cannot reach. | No. Google states no third-party tool has access to its ranking or AI systems, and that includes ours. |
The standard we will be held to
What has to be in a published result before it means anything?
Five things. A number without them is a number the vendor chose, and there is no way for you to tell the difference from outside.
- The baseline figure, and the date it was taken.Without a starting number, a rise has no size. "Up 40%" from what?
- The calendar date range.Model versions roll unannounced. A result with no dates cannot be placed against the week an engine changed underneath it.
- The question list, in full.A visibility total summed across questions grows when you add questions. If the list is withheld, the total is unbounded and cannot be reproduced.
- The measurement tool, and its version.Every vendor score in this category is an outside estimate — Google says so. Which estimate, computed how, is the whole question.
- The client, on the record.First-person agency voice is not verification. A named person at the brand, attesting to the figure, is.
We read the AEO case studies we could find published in this category in September 2026. None disclosed all five. SearchD has published none either — we have no results yet. When we do, they carry all five or they do not go up, and you can hold us to that sentence.
Common questions
Is AEO just SEO with a new name?
For Google’s surfaces, Google says yes: "optimizing for generative AI search is optimizing for the search experience, and thus still SEO." That settles one surface and leaves the others open — ChatGPT, Perplexity and Claude retrieval is not addressed by that guidance, and that is where work an SEO shop cannot claim actually sits. SearchD treats the Google surface as SEO hygiene we verify, and spends the engagement on the surfaces it does not cover.
If the evidence is this thin, why hire anyone?
Because the diagnosis does not depend on the contested research. Whether an engine can fetch your site, whether your specification exists as text a model can lift, whether an assistant is stating a wrong SPF value about your product — those are first-party observations, and you can verify any of them yourself in a browser in under a minute. The contested part is the causal chain from a fix to a citation, and that is exactly the part SearchD measures rather than promises.
How is AI referral traffic doing in absolute terms?
Small, and worth saying out loud. Independent clickstream panels put all AI referral traffic well under one percent of total web visits, and the largest single destination of ChatGPT’s outbound links is Google. SearchD has not reproduced those panel figures itself, so it does not quote a precise number for them here. The case for the work rests on where buyer research is moving, not on the current referral volume.
See who US buyers find instead of you.
Send us your URL. We run 20 buyer questions across five engines by hand and send you the transcript back within 48 hours.