AI search visibility for e-commerce & SaaS
I measure how AI answer engines treat your site — and fix what is provably in your control.
Nobody controls whether ChatGPT picks you for an answer. I watched a store's ChatGPT traffic appear to collapse overnight, showed from server logs that most of it was lost tracking, and that the real change — a fourfold fall in ChatGPT's own fetches — was not caused by the site. What can be measured is measured here; what can be changed is changed with a before-and-after you can check yourself.
- same day delivery [category] dubai
- who ships [category] to abu dhabi
- cash on delivery [category] uae
The model answers these from what your own site states. Publish the fact clearly and it can be quoted — when it is retrieved, and whether you are the one picked, the engine decides.
- best [category] brand
- most trusted [category] store
- is [brand] any good
Here the model ignores your claims and reaches for third-party evidence — reviews, forums, editorial. No amount of on-site copy moves these.
Across twenty tracked purchase queries this split held without a single exception[1]. It is the first thing I check on any brand, because it tells you which half of the problem sits on your own site and which half needs a different instrument entirely.
What is sold here
The measurable part, and nothing else.
- A measurement, not a screenshot. Your buyers' questions, three runs each across ChatGPT and Perplexity, in one fixed mode, with a control question. Sixty recorded answers, every source listed.
- Google AI Overviews from first-party data. Search Console reports how often your pages appear in Google's AI features. Most site owners have never opened that report.
- Crawler telemetry from your own server. Which AI crawlers visit, what they take, what your server answers them — and whether ChatGPT's live fetcher is still reaching you, independent of referrer tracking.
- Technical fixes with a measurable cause and effect. One change at a time, verified in the logs the next day.
- The negative findings. Which questions in your category cannot be won from your site, written down before you pay for anything.
What I do not sell
- Being cited by ChatGPT. Selection happens inside the engine; I can show you when it changes, not make it change.
- Positions or rankings. They do not exist in answer engines.
- Timelines. OpenAI's search crawler can take weeks to reach new pages.
The case study
Eight weeks, one bootstrapped store, no ad spend.
I own and run a specialty e-commerce brand in the UAE. No paid acquisition, no team, competing against businesses with far more capital. That made AI answer engines the only channel worth committing to — so I built the playbook on my own money before selling it to anyone.
Tracked purchase queries returning the brand in the top five of ChatGPT and Perplexity, from zero at baseline[1]. An observation over eight weeks, not a proof.
First-named source on the highest-intent question in the category, at the time of the measurement[2].
ChatGPT reproduces the site's own phrasing inside its answers. Not just cited — described in our words[3].
Control queries behaved as predicted, so the movement was not model drift. What caused it is not isolated[1].
Read these as observations, not as proof of what the work did. There was no control group: the domain aged, competitors moved, and many things changed at once. In October ChatGPT's live fetches of the same store fell fourfold in three days, with no change to the site that could explain it — the engine moved, not the site. The one result with clean cause and effect was technical: replacing 439 placeholder pages with real redirects cut Googlebot's requests to dead addresses from 23 to 3 a day, visible in the logs within a day.
And the part that failed
One sprint produced thirty purpose-built landing pages aimed at “where to buy” queries. Eleven days later they had earned zero citations — not few, zero. The answer engines were resolving those questions from pages that already existed and never touched the dedicated ones[4].
I mention it because it is the most useful thing I learned all summer, and because a consultant who only shows you the winning sprints is showing you half a method.
Method
Four layers, deployed in order.
The sequence matters more than any individual piece. Each layer only pays off once the one beneath it is in place.
Retrieval
Structured data, a crawlable homepage and an llms.txt manifest, all written as machine-readable fact rather than marketing copy. What is measurable: published phrasing comes back verbatim once a page is retrieved — and when that happens is not under the site's control. Which file actually carries it is a separate question, and my own logs do not settle it in llms.txt's favour[5].
Answer surface
The pages that actually get quoted — homepage and category pages rebuilt to answer the question directly in the first screen, in the words buyers use rather than the words the brand prefers.
Query-shape coverage
Content generated against the real question patterns for your category and markets, then tested under a dual-framing protocol because research-phrased and purchase-phrased versions of the same question return different answers.
Third-party evidence
The layer that unlocks judgment queries, and the slowest of the four. Reviews, independent mentions, comparison surfaces. Nothing on your own domain substitutes for it — which is exactly why most brands stall here.
Working together
Everything is delivered in writing.
No discovery call required, no standing meetings. You get a written document and a recorded walkthrough you can read at 11pm and forward to whoever needs it. Most founders prefer this; it also means the reasoning is on paper instead of in someone's memory of a call.
Four of your buyers' questions in ChatGPT and Perplexity. The answers verbatim, every source listed. No scoring, no pitch.
Is your AI traffic real? From fourteen days of server logs: which pages ChatGPT and Perplexity actually read, whether a drop is lost tracking or lost visits, and what your structured data claims that your pages do not. No prompts.
Where should the money go? Sixty recorded answers reduced to the two to four sites that decide your category, whether you are on them, what each requires — and which popular tactics do not work in your category, before you pay for them.
Getting onto those sites, fixing what the logs showed, and measuring every month. What is guaranteed is the work and the measurement. Nobody can guarantee a citation, and I will not.
Start here
Free category check.
Tell me the brand and the questions your buyers ask. I run four of them by hand in ChatGPT and Perplexity and send back the answers verbatim, with every source listed under its answer. No score, no pitch, nothing scheduled afterwards unless you ask.
Provenance
How these numbers were measured.
Figures on this page come from one documented measurement run, not from a dashboard estimate. Where a number would not survive scrutiny, it is not here.
- [1] Structured measurement of twenty purchase-intent queries across ChatGPT and Perplexity, run against a baseline recorded before deployment, with four control queries to separate real movement from model drift. Same session protocol and query wording on both runs.
- [2] First-named source in the answer body, observed on repeated runs of the trust query in the brand's own category and market. Screenshots retained.
- [3] Phrases published on the site reappearing word-for-word inside answer text, across multiple query variants — the difference between being listed and being described. Which page the model lifted them from is not something my logs establish; see note 5.
- [4] Thirty template-generated URLs, live and indexed, checked eleven days after deployment: no citations on any tracked query. Reported here because the negative result changed the method.
- [5] On llms.txt specifically, because the industry treats it as settled and I no longer do: across fourteen days of my own server logs, the only named crawlers to request /llms.txt were Bing's and Yandex's. OpenAI's, Anthropic's, Perplexity's and Google's crawlers fetched the homepage and robots.txt and skipped the manifest. The verbatim-phrasing effect is measured and real. The claim that llms.txt is the channel carrying it is not something I can currently evidence, so I have stopped making it. I still publish one — it costs nothing and Bing does read it.
- NDA The brand is a live business and stays confidential on public pages. Traffic dashboards, order attribution and raw answer-engine screenshots are available for direct verification under a mutual NDA once we are talking seriously.
Measurements
ChatGPT traffic collapsed overnight. Most of it was lost tracking — one number was not.
In October a store's tagged visits from ChatGPT fell from about 23 a day to 2–3, while direct visits rose by 15 a day on the same landing page — together, a fall of about 15%. What no tracking change can hide, ChatGPT-User fetches in the server log, fell from 26.5 a day to 7, while every other crawler carried on. Daily tables, a correction to the first version, and the commands to repeat it on your own server.
The cut-off is not quoted. It is calculated with.
On 21 September 2026 an answer engine read a retailer's published delivery cut-off, compared it against the current hour in the buyer's city, and demoted the category leader because the window had closed. Reproduced independently on two accounts. Full method, exact queries and verbatim answers.
Skincare in the UAE: the label held, the winner did not.
Four buyer questions about skincare and body care, ChatGPT and Perplexity, two dates five days apart, 24 recorded answers. The label an engine put on a brand repeated in 6 of 6 answers while the first pick changed; four sites built most of the answers — three pharmacy and retail storefronts and one newspaper. Reddit, Quora and LinkedIn: zero.
And a second one, on this category itself: twenty recorded runs of which AI-visibility agencies the engines actually name, how often, and what they cite when describing them — including the measurement where this site scored zero out of twenty.
Services
Two questions I answer, and one job I do
Start with the free category check: four of your buyers' questions across two engines, answered in two working days. If your category returns no vendor list to anyone, you will be told that instead of sold something.
Since October ChatGPT passes its tag on fewer clicks, so dashboards show collapses that did not happen — and miss ones that did. Fourteen days of your server logs settle it: what the engines read, what was really lost, what your structured data gets wrong. No prompts. Three working days, $400.
In most categories two to four third-party sites decide what AI answers say. Sixty recorded answers find them; you learn whether you are on them, what each requires, and which tactics do not work in your category. Five working days, $600.
Then, if you want it done: the placement program — getting onto those sites, fixing what the logs showed, measuring every month. Three months minimum, from $1,500 a month.
Selling to consumers? The same work for an online store, with the measured findings that apply only to retailers — delivery windows, stock status, and the two formats that were measured and failed: GEO for e-commerce →