A meaningful share of the questions your customers used to type into Google are now asked of an assistant instead, and the answer they get names two or three products. If you are not among them, the visit does not happen — and unlike a poor Google ranking, there is no report telling you.
The instinct is to assume good SEO covers it. The research does not support that assumption: studies of AI-search citation consistently find that the overlap between what ranks organically and what gets cited is much lower than practitioners expect, and that citation behaviour differs dramatically between assistants. Ranking first is helpful and it is not sufficient.
A note on the numbers in this area. Most published statistics come from tool vendors with an obvious interest, methodologies are rarely disclosed, and the platforms change monthly. Treat every figure you read — including the direction of the ones here — as indicative, and measure your own position rather than trusting a benchmark.
Why the Two Systems Diverge
Search ranks documents against a query. An assistant composes an answer, and to do that it needs sources it can extract specific, verifiable statements from. Those are different jobs and they reward different content.
- Extractability over persuasion. A page that states a specification plainly is more usable than one that argues beautifully around it. Marketing prose is hard to quote.
- Specificity over breadth. A comprehensive category guide competes with everyone. A page that definitively answers one narrow question is quotable.
- Structure over volume. Clear headings, direct answers, tables and lists survive extraction. Long undifferentiated prose does not.
- Corroboration matters. A claim repeated consistently across your site, your product data and third-party sources is safer for an assistant to repeat than one that appears only in your marketing copy.
- Freshness is weighted differently. For anything with a date-sensitive answer, currency counts for more than accumulated authority.
What Actually Earns a Mention
1. Product Data a Machine Can Trust
This is the foundation and it is where most stores fail before anything else matters. An assistant asked for waterproof walking boots under £150 in a wide fit can only consider products whose waterproofing, price and fit are stated in a form it can read.
Attribute completeness, controlled vocabularies, explicit units, GTIN and brand — the whole unglamorous catalogue project. We cover the enrichment method in cleaning a catalogue with AI and the scoring in the 40-point agent-readiness audit.
2. Structured Data That Agrees With the Page
Product, Offer, availability, returns policy and shipping as validated schema, matching what a human sees. The contradiction case matters more here than in conventional SEO: a price in markup that differs from the price on screen does not merely fail to help, it makes everything else you publish less trustworthy.
3. Answers to the Questions People Actually Ask
Assistants are asked comparative and conditional questions — which of these is better for X, does this work with Y, what should I buy if Z. Most commerce sites publish nothing that answers those, because the category page is a grid and the product page is a description.
The content that gets cited is the buying guide that genuinely compares, the compatibility page that states what fits what, and the honest this product is not right for you if paragraph. That last one is disproportionately effective, because it is exactly the kind of qualified statement an assistant can safely repeat — and almost nobody publishes it.
4. Being Corroborated Elsewhere
Assistants weight consistency across sources. Consistent product data in marketplace listings and merchant feeds, accurate business information wherever it appears, and genuine third-party coverage all make your own claims safer to repeat. This is unfashionable, slow work and it is the part with the longest half-life.
5. Being Reachable
The failure that costs the most and takes the least effort to fix. Check your logs for the assistant crawlers you care about and confirm they are getting content rather than a challenge page. Aggressive bot protection, JavaScript-dependent rendering and blanket crawler blocks all quietly remove you from consideration, and none of them announce it.
Before any content work, spend an hour in your server logs. We regularly find stores investing in visibility while returning challenge pages to the exact crawlers they are trying to reach.
What Does Not Work
- Keyword stuffing for assistants. They are not matching keywords; they are extracting claims.
- Publishing volume. Twenty thin comparison pages are less citable than one genuinely thorough guide.
- Prompt injection in page content. Instructions hidden in markup aimed at manipulating an assistant. It does not reliably work, it is trivially detectable, and being caught is a reputational cost with no upside.
- Chasing the acronym. The industry cannot agree whether this is GEO, AEO or AI visibility, and the search volume for all of them is small. Target the customer's actual question, not the discipline's name.
- Buying a monitoring tool before doing the data work. You will pay monthly to watch a number that cannot move.
Measuring It Without Buying Anything
The tooling market is real and some of it is good, but you can establish a baseline yourself in an afternoon.
- Write a fixed panel of 50 queries a real customer might ask an assistant in your category. Mix broad (best X for Y) with narrow (does A work with B). Fix the wording and do not change it — the panel's value is that it is comparable over time.
- Run it monthly across the assistants your customers use. Record: do you appear, in what position, how are you described, and who appears instead.
- Note what is being said about you, not just whether you appear. A citation that describes your product wrongly is a problem to fix at the source.
- Track AI referral traffic separately in analytics, with its conversion rate. It is frequently higher than non-branded organic, which is the argument that funds the work.
- Watch your logs for assistant crawler activity, which tells you what is actually reaching you regardless of what any dashboard reports.
- Review quarterly, not weekly. These systems change too often for weekly readings to be signal.
Keep the panel and the results in one place from the start. The single most useful artefact here is a year of comparable monthly readings, and it only exists if you begin recording before you have anything to show.
An Honest Assessment of the Opportunity
For most retailers today, AI-referred traffic is a small share of sessions and a growing one, with conversion rates that tend to be favourable because the intent is qualified before the click arrives. It is not yet a channel that replaces search, and anyone telling you otherwise is selling something.
What makes it worth acting on now is that the work is not speculative. Attribute completeness, structured data, genuine buying guides and a sane crawler policy improve conventional search, on-site search and merchant feeds at the same time. You are not betting on the channel — you are doing overdue work that happens to also serve it.
The part that is speculative — agentic checkout, protocol adoption, transaction interfaces — can wait, and we say so in the agent-readiness audit and in your product needs an MCP server.
We do this work through our AI integration practice, with the catalogue and engineering side handled by our AI and machine learning team and the content and measurement side by our digital marketing team. For retail brands it usually starts, unglamorously, with the attribute audit.
Frequently asked questions
Does ranking on Google mean AI assistants will cite us?
Much less than people assume. Studies of AI-search citation consistently find limited overlap between what ranks organically and what gets cited, and citation behaviour varies enormously between assistants. Good SEO helps; it does not cover this.
What is GEO, and should we care about the term?
It stands for generative engine optimisation, and the industry has not settled on it — AEO and AI visibility describe the same work. Search volume for all of them is small. Target the customer's actual question rather than the discipline's name, and treat the acronym as a label for the work, not a strategy.
How do we measure AI visibility without buying a tool?
Build a fixed panel of 50 realistic customer queries, run it monthly across the assistants your customers use, and record whether you appear and how you are described. Track AI referral traffic separately in analytics and check your logs for assistant crawlers. Paid tools save time; they are not a prerequisite.
Can we optimise content specifically for AI assistants?
Yes, and it looks like good content rather than a trick: direct answers to specific questions, comparisons that genuinely compare, compatibility stated explicitly, and honest statements about who a product is not for. Extractable, verifiable claims are what get quoted. Hidden instructions aimed at manipulating assistants do not work and carry reputational risk.
How much traffic does AI search actually send?
For most retailers it is a small but growing share of sessions, with conversion rates that tend to be better than non-branded organic because the intent is qualified before the click. It is not replacing search yet. The reason to act is that the underlying work improves conventional search and on-site search at the same time.
Start With the Panel and the Logs
Two hours of work gives you a baseline and tells you whether you are even reachable. We will run both against your store and tell you what is actually blocking you. Talk to our team; we reply within a business day.