Reach · Getting recommended by ChatGPT, Gemini & Perplexity
Be named in the answer, not ranked below it.
You show up in Google and nobody at the brand can say whether ChatGPT recommends you. We build the data, the pages and the third-party citations that decide it — and we measure whether your name actually appears.
-
8×
growth in AI-driven traffic to Shopify stores, year over year
-
≈85%
of brand mentions in AI answers come from pages you do not own
-
+31%
conversion rate on ChatGPT referral traffic, against non-branded organic
Sources: Shopify president Harley Finkelstein, Q1 2026 earnings call; AI-search citation research; ChatGPT referral conversion measured across 94 ecommerce brands. Full table with attribution below.
The problem
A $6M brand on Shopify Plus, billing subscriptions through Recharge, shipping from a 3PL and selling the same fourteen SKUs on Amazon. Google is fine — Search Console says non-branded impressions are flat, which is what they were last quarter. The question nobody in the building can answer is the other one: when a buyer opens ChatGPT and types “best magnesium glycinate for sleep” or “which creatine actually dissolves,” does the answer contain the brand’s name? There is no export for that. There is no impression count. There is one person who tried it on their phone during a meeting, got three competitors, and never brought it up again.
This is not a rounding error any more. Shopify’s president Harley Finkelstein told the Q1 2026 earnings call that AI-driven traffic to Shopify stores grew 8× year over year and orders from AI-powered search grew nearly 13×. Across retail more broadly, AI-driven traffic rose 693% during 2025. And the traffic is better than the traffic it is taking share from: measured across 94 ecommerce brands, ChatGPT referrals converted at 1.81% against 1.39% for non-branded organic — 31% higher — while revenue per AI-referred session ran 10.3% above organic. Small base, steep curve, better visitors. That is the shape of a channel worth being early in.
The reflex is to hand it to whoever owns SEO, and that reflex is the expensive mistake. Classic search optimisation is a first-party game: you control the page, the page ranks, the click arrives. AI answers are largely a third-party game. Roughly 85% of brand mentions in AI answers originate on pages the brand does not own — review platforms, comparison posts, category round-ups, marketplace listings, forum threads. You can rewrite every page on your own domain and remain unnamed, which is not a sentence anybody could say about a Google ranking.
So the work splits, and not the way the budget usually does. The part on your own site is real, but it is smaller than anyone expects and it is not what an SEO brief would tell you to do there. The larger part sits on pages you do not control, and it looks more like catalogue hygiene and public relations than like keyword work.
The duller reason brands go unnamed
A retrieval system has to be confident about what your brand is before it will put your name into a sentence it is accountable for. If your Shopify product title, your Amazon listing, your Google Merchant feed and your review platform each describe the same SKU differently — a different count, a different flavour name, the brand written two ways, a strength that changed last spring on one channel and not the others — you have handed it several contradictory descriptions of one object. Contradictions do not get resolved. They get skipped, in favour of a competitor whose data agrees with itself.
Answer-shaped, not article-shaped
The on-site half is a rewrite, not an expansion. A page that answers a real question in its first sentence, states the specification in an HTML table and cites where the claim comes from is a page an engine can lift a passage out of. A 2,000-word post that opens with the founder’s origin story and puts the dosage in an image at the bottom is not. A Princeton study found that these techniques — citing sources, adding statistics, restructuring so a claim and its evidence sit together — lifted a page’s visibility in AI responses by 30–40%. That is a real finding about page-level visibility under controlled conditions. It is not a revenue number at a brand, and nobody has published one of those worth believing yet. We say the same thing on the phone.
Where this sits in the market
Our own read of the category in 2026, not a third-party index — we have an obvious interest in calling this early and uncrowded, so treat it as a point of view rather than a finding.
What we build
Nine pieces. About half sit on your own property and about half do not, which is the shape of the channel rather than a preference of ours. Everything is built in your accounts and documented as it goes.
Engines we build and measure against
- Structured data on real templates Product, Offer, Organization, Review and FAQPage schema written into the theme rather than injected by an app that a theme update will silently break. Correct prices, correct availability, correct variant-level attributes, validated against the live pages instead of a staging copy.
- One canonical entity, everywhere The brand name spelled one way, the product names spelled one way, the same founding facts and the same category description on Shopify, Amazon, TikTok Shop, your review platform, your Google Business profile and every publisher listing we can reach. A retrieval system that finds three descriptions of one company cites none of them.
- Clean feeds across channels One canonical product record mapped out to the Google Merchant feed, Amazon, TikTok Shop and Walmart so the size, the flavour, the count, the ingredient list and the price agree everywhere. On most accounts this has to happen before anything else on this page is worth doing.
- llms.txt, published honestly A plain-text map at the root of the domain saying what the site contains and which pages are the canonical answer for what. It costs an afternoon and a static file. Read the section below before you decide what it is worth — we publish it because it is cheap, not because it is proven.
- The third-party surface The review platforms, comparison posts, category round-ups, marketplace listings and forum threads that the engines actually cite for your prompts. We pull the cited URLs out of the answers themselves, then work the sources that can legitimately be worked: correcting an outdated spec, submitting to round-ups that accept submissions, making the marketplace listing agree with the site, answering in the threads where your buyers already are, under your own name.
- Answer-shaped pages on your own site Real questions as headings, the answer in the first sentence beneath — not in paragraph four after the brand story. Specifications, sizes, ingredients and compatibility in real HTML tables, not an image, not a tab widget, not a PDF. A model lifts a table. It does not lift a screenshot.
- Comparison and alternatives pages The head-to-head and “best X for Y” pages most brands will not write, including the honest paragraph about who you are not for. They get cited because they contain a comparison a model can quote, and they are one of the few third-party-shaped assets you can own outright.
- The prompt set, run on a schedule Forty to eighty prompts your buyers actually type, run across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews on a fixed cadence, logged three ways: named, not named, or named alongside which competitors. Every cited source captured, because that list is the work plan for the following month.
- AI referral tracking that survives AI referrers separated in GA4 and in your warehouse so the traffic is not sitting in Direct. Then the part volume dashboards never show: conversion rate, AOV, subscription attach rate and refund rate on AI-referred sessions against every other source.
The third piece is the same work as catalogue and feeds, and on most accounts it has to happen first. The ninth lands in reporting and analytics, because a visibility number with no revenue beside it is half an answer.
How the build runs
Six to eight weeks from access to a measured baseline and a working prompt set. The on-site half finishes first because it is the half you control.
-
01
Prompt set & baseline
We build the prompt set with your team — the questions a buyer asks before they know your name — and run it across all five engines. You get a baseline that says how often you are named, which competitors are named instead, and exactly which URLs the engines pulled from. Nothing gets built before that list exists.
Week 1
-
02
Entity & data layer
Structured data written into the theme, canonical naming reconciled across every property, feeds corrected channel by channel, llms.txt published. Unglamorous, mostly invisible, and the reason anything downstream is trusted.
Week 1–3
-
03
Answer-shaped rebuild
The pages that should answer the prompt set, rebuilt to answer it: question headings, answer-first paragraphs, real tables, comparison and alternatives pages written where the prompt set says they are missing.
Week 2–5
-
04
The third-party surface
Working the sources the baseline showed the engines citing. Corrections and updates first, because they are fastest and they are yours by right. Then submissions, listings and threads. This runs on other people’s editorial timelines, which is why it is the longest bar on the chart.
Week 3–8
-
05
Measurement
The prompt set re-runs on a schedule, the AI referrers get split out in GA4 and the warehouse, and the dashboard shows both halves: whether you are being named, and what the resulting traffic does after it lands.
Week 4
-
06
Re-run & iterate
Monthly re-runs against the baseline. The engines change their retrieval behaviour without announcing it and without a changelog, so the only defensible position is measuring your own prompt set often enough to notice when something moved.
Ongoing
The evidence, with attribution
Every figure this page argues from, with who said it. There is no benchmark table of our own results here yet, and the last row says so rather than filling the gap with something invented.
| What it measures | Figure | Attribution |
|---|---|---|
| AI-driven traffic to Shopify stores, year over year | 8× | Harley Finkelstein, Shopify president — Q1 2026 earnings call |
| Orders from AI-powered search on Shopify, year over year | ≈13× | Harley Finkelstein, Shopify president — Q1 2026 earnings call |
| ChatGPT referral conversion rate against non-branded organic | 1.81% vs 1.39% | Measured across 94 ecommerce brands |
| AI-driven traffic to retail sites, 2025 | +693% YoY | Retail-sector AI traffic study, 2025 |
| Revenue per AI-referred session against organic | +10.3% | Retail-sector AI traffic study, 2025 |
| Brand mentions in AI answers originating on third-party pages | ≈85% | AI-search citation research |
| Visibility lift in AI responses from GEO techniques | 30–40% | Princeton GEO study |
| Prompt-set win rate across our own client builds | — | Pointerflow client data — metric to confirm |
Where we can name the speaker we do — the two Shopify figures are Harley Finkelstein’s own words on the Q1 2026 earnings call, which makes them a public company’s stated numbers rather than an independent audit. The remaining third-party figures are published measurements we have not audited, each measuring a channel that is moving month to month, and none of them is a forecast for your account. The Princeton finding is the most rigorous thing published in this category and it measures page-level visibility under controlled conditions, not revenue at a brand. The final row stays an em dash until we have enough client builds behind it to publish something honest.
What is not settled yet
Three things in this category are being sold as finished. None of them are. You will be pitched all three, so here is where each actually stands.
Agentic checkout was pulled back, not shipped
OpenAI launched Instant Checkout in ChatGPT in September 2025 and withdrew it on 4 March 2026, with only around 30 merchants live at the end. Walmart — one of the retailers with the volume to measure it properly — found in-chat checkout converting roughly 3× worse than a click through to walmart.com. So the pattern that works today is discover in AI, buy on site, and that is the pattern we build for. We do not sell agentic-checkout readiness as a finished product, because the product was withdrawn six months ago.
UCP is the more complete standard, and it is still moving
Google and Shopify’s Universal Commerce Protocol is a more serious attempt at the same problem, and it is not finished either. We track it and we do not pre-sell it. What is worth noticing is that the groundwork is identical: a clean canonical product record, consistent entity naming and structured data that is actually correct are what a protocol would read, and they are also what gets you named in an answer today. That is the only kind of future-proofing worth charging for — work that pays whether or not the standard lands.
llms.txt is a proposal, not a standard
It is a proposed convention with partial support. Perplexity has stated support. Google, OpenAI and Anthropic have not formally committed to reading it. We publish it anyway, because it is an afternoon and a static file and the expected value is positive even at a low probability of it mattering — but that is the whole argument for it, and it is an argument from cost rather than from evidence. If somebody prices llms.txt as a deliverable in its own right, you now know what you are being sold.
There is a fourth, which is about us rather than the category. The engines change their retrieval behaviour without a changelog and without telling anyone. A visibility position is not a thing you win once; it is a thing you re-measure. Any agency describing this as a project with an end date has either not run the prompt set twice, or is not telling you what happened the second time.
Prompt-set win rate across our own client builds
—
Metric to confirm. We will publish it when there are enough builds behind it to publish something honest. Meanwhile, treat every “GEO lift” figure you are quoted in this category with the same suspicion: most are vendor-reported, on the vendor’s own customers and the vendor’s own definition of a mention. Independent measurement here is thin, the Princeton study is the most serious published work and it measures page visibility rather than revenue, and the engines it was run against have changed since. See reporting and analytics for how we measure your own account instead.
What it costs
Published ranges, because hidden pricing costs more leads than it protects. Where you land inside the range is decided by how many channels your product data lives on and how much of it currently disagrees — not by what we think you can pay.
Revenue Recovery Audit — the visibility baseline sits inside it
$1,500–$3,000
AI search visibility build
$4,000–$12,000
Ongoing — prompt-set re-runs, third-party work, new answer pages
from $2,000/mo
Feed and catalogue remediation, where the product record is the blocker
Quoted with the feed work
The audit is credited in full against any build you go ahead with, and the visibility baseline — your prompt set, run once across all five engines, with the cited sources listed — is part of it. That baseline is useful whether or not you hire us for the build, and it is the honest way to find out whether this is your biggest leak or your fifth. If the audit says your catalogue is the blocker, we will tell you to spend the money on the feed first.
Where to go next
Questions
Is this just SEO with a new name?
No, and treating it that way is the expensive mistake in this category. Classic search optimisation is a first-party game: you control the page, the page ranks, the click arrives. AI answers are largely a third-party game — roughly 85% of brand mentions in AI answers originate on pages the brand does not own. The overlap is real, because crawlable HTML, correct structured data and a fast site help both. But you can rewrite every page on your own domain and still not be named, which is not a sentence anyone could say about Google rankings.
How do you measure whether it is working?
Two ways, and both are needed. First, a fixed prompt set — forty to eighty questions your buyers actually type — run on a schedule across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, logged as named, not named, or named alongside which competitors, with every cited URL captured. Second, AI referrers split out in GA4 and your warehouse so you can see conversion rate, AOV, subscription attach and refund rate on that traffic rather than only its volume. A visibility report with no revenue side to it is half an answer.
Should we publish llms.txt?
Yes, but for the right reason. It is a proposed convention with partial support: Perplexity has stated support, and Google, OpenAI and Anthropic have not formally committed to reading it. It costs an afternoon and a static file, so the expected value is positive even at a low probability of it mattering. Publish it because it is cheap, not because it is proven. If somebody quotes you a line item for llms.txt as a deliverable in itself, that is the tell.
Should we be getting ready to sell inside ChatGPT?
Not as a priority, and be careful who tells you otherwise. OpenAI launched Instant Checkout in ChatGPT in September 2025 and pulled it back on 4 March 2026 with only around 30 merchants live; Walmart measured in-chat checkout converting roughly 3× worse than a click through to walmart.com. Google and Shopify’s Universal Commerce Protocol is the more complete standard and is still moving. The pattern that works today is discover in AI, buy on site, and that is what we build for. The groundwork that would make you ready for agentic checkout later — a clean canonical product record, consistent entity naming, structured data that is actually correct — is the same work either way, which is why we sell that and not the speculation.
How long before we get named?
The first re-run of the prompt set is at week four and it usually shows movement on the on-site half, because those changes are yours to make and the engines re-crawl. The third-party half moves over a quarter, not a fortnight, because corrections, listings and round-ups run on other people’s editorial timelines. Anyone promising you a placement in an AI answer by week two is describing either a paid placement or a guess.
Our catalogue disagrees with itself across Amazon and TikTok Shop. Does that matter here?
It is the blocker, and it is the most common one. A retrieval system reading three different titles, two different sizes and a fourth ingredient list for the same SKU has no reason to trust any of them, and contradictions get skipped rather than resolved. On those accounts we fix the canonical product record first — that is the catalogue and feeds work — and do this on top of it. We will say so in the audit rather than sell you the visibility build anyway.
Does this only apply to subscription brands?
No. Nothing on this page depends on billing to a schedule. What it depends on is selling something people research before they buy, and a category where buyers ask questions with a shape — which fits a supplement brand, a skincare line, a home-goods catalogue, a specialty food brand and most of what a $3M–$30M ecommerce operator sells. Subscription brands get one extra advantage from it: an AI-referred subscriber is acquired once and billed repeatedly, so the higher conversion rate on that traffic compounds instead of ending at the first order.
Does the third-party work mean buying links?
No. What it means is reading the cited URLs out of the answers themselves and then working the ones that can legitimately be worked — correcting an out-of-date specification on a review platform, submitting to category round-ups that accept submissions, making a marketplace listing agree with the site, answering questions in forums under your own name and with the affiliation stated. Where the honest answer for a source is that you cannot influence it, we say that instead of billing for the attempt.
Find out what you’re losing.
Before you commit to anything, we tell you exactly what you’re losing and what it costs to stop it. Two weeks. Fixed fee. Credited in full against any build you go ahead with.
- Fee
- $1,500–$3,000, fixed
- Duration
- Two weeks
- Credited
- In full, against any build
- You supply
- Read access + one 45-minute call