Shopify published a year of category-level data on AI-referred traffic. The growth number is the least interesting part of it.
On August 11, Shopify's enterprise blog published a year of category-level data on how AI search refers shoppers to its merchants. Kyle Risley wrote it up at https://www.shopify.com/enterprise/blog/ai-search-category-behavior and it is worth reading in full. The headline is the number everyone will quote: AI-referred sessions to Shopify storefronts grew 197% year over year, and orders grew about 3x.
That is also the least useful number in the piece. Shopify is careful to say that organic search still referred more sessions than every AI platform it tracks put together, and that organic grew 12% on a much larger base. The correct read is not that search has been replaced. It is that a small, fast-growing channel behaves differently enough from the big one that averaging them together hides what is actually happening.
Three findings point the same direction. About half of AI-referred sessions landed directly on a product page rather than a home page or a collection. Spec-heavy categories showed the widest gaps: watches converted about 2.4x better from AI-referred traffic than from organic, necklaces about 2.3x, apparel overall about 1.6x. And AI search introduced net-new customers at roughly 1.3x the rate of organic.
Together those describe a visitor who has already done the comparison somewhere you cannot see. The engine did the shortlisting, the category education, and often the spec matching. What lands on your site is the end of a research process rather than the beginning of one. Shopify also found that when an engine had structured catalog data to work from, the shoppers it sent converted at twice the rate of those from sessions relying on scraped or third-party feeds, which is a polite way of saying that a model's confidence in your product facts shows up in your conversion rate.
Worth noting what the post does not include: no sample sizes, no margin of error, no methodology. This is first-party platform data, directionally useful, not a controlled study. Read it as a strong signal rather than as numbers you can port onto your own forecast.
The first is classification. If AI-referred visitors land deeper and convert differently, then bundling them into organic makes both groups read wrong at once: your organic conversion rate is flattered by traffic that was never organic, and the AI cohort is too small a slice of the blend to argue with. You cannot act on a channel you have not separated.
The second is harder, because it happens before any traffic exists. An answer engine has already described your category to a buyer, named some companies, and cited some sources. If you were not among them, there is no session to analyze. That absence never appears in analytics, because nothing happened. This is the part of answer engine optimization that ordinary reporting structurally cannot reach.
Two things, kept deliberately separate. AI-referred traffic detection ships with the tag and needs nothing else switched on. AXO classifies arrivals from ChatGPT, Perplexity, Gemini, Copilot, and Claude by referrer and campaign source, and keeps that as its own acquisition axis rather than folding it into organic, so your existing organic segments keep meaning what they always meant. It becomes a targetable audience like any other, and no PII required.
One deliberate omission there is worth stating, because it cuts against our own numbers: Google AI Overviews is not counted. An AI Overview referral arrives as plain google.com and is indistinguishable from ordinary search. Counting it would mean misclassifying regular organic traffic to inflate an AI figure, so AXO does not count it.
The AI Visibility module addresses the second blind spot. You give it the questions your buyers actually ask, and it puts them to answer engines: Perplexity, Claude with web search, OpenAI with web search, grounded Gemini, and Google AI Overviews through a SERP provider. For each answer it records whether you were named at all, where you fell in the order of companies mentioned, the sentence used to describe you, its tone, whether the claims made about you were factually correct, every competitor named alongside you, and every source the engine cited. Results are reported per engine rather than averaged, because being invisible on a single engine is precisely the finding an average erases.
Measuring is the first half. The module localizes a gap precisely: this question, this engine, this competitor took the slot, this third-party page got cited instead of yours. Closing it is the second half, and it runs as its own loop on top of the same data.
AXO turns a run into findings rather than a wall of results. A finding is one specific, deduped, ranked problem with its evidence attached: a question you lose on more than one engine across more than one run, a directory the engines keep citing where you have no listing, a claim about you that came back factually wrong. Alongside the engine results it audits your own site for the things that make you unreadable in the first place, and the most common one is also the cheapest to fix. A robots.txt that quietly blocks GPTBot or ClaudeBot means you cannot appear in those answers at all, no matter how good the content behind it is.
For most findings AXO then generates the artifact that closes it. Schema.org markup for your organization, your services, and your products. FAQ blocks that answer the questions you are losing, in the words your own pages already use. Product attribute enrichment for the spec-level facts an engine wants before it will recommend you. Patches for robots.txt and llms.txt. One rule governs everything structured: it is assembled from your own data, never written by a model. When a required fact is missing, AXO asks you for it instead of inventing a value to fill the field. Schema that asserts something untrue is a machine-readable lie told at scale in your name, and no amount of convenience is worth that.
Then it stops and waits for you. Every artifact lands as a draft. You read it, you approve it, and AXO ships it. Nothing publishes itself, and there is no setting that changes that. This is not caution for its own sake: these artifacts describe your business to five models at once, and the cost of a wrong one is that the models now believe it.
One thing we are careful about, because the category generally is not. Most AI crawlers do not execute JavaScript, so anything a tag injects into a page after load is invisible to them. AXO labels which route reaches which audience instead of selling blanket coverage: on-page content published through the tag reaches your visitors and Google, and structured data has to land in your site's own HTML to reach the rest. You will always be told which one you are getting.
A few limits worth stating plainly. AI Visibility is an opt-in module on Growth plans and above, off by default, because every run spends real money at the engine APIs. Runs are manual unless you set a cadence, so nothing bills you by surprise. Engines without a configured API key are skipped visibly and recorded on the run rather than quietly dropped, so a narrower run always reads as narrower instead of looking like a worse result. The structured artifacts are assembled from data you already have and cost nothing beyond the run itself; only the written blocks and briefs draw on AI credits, and those are capped per run.
And it does not decide for you. Approval is a person reading the thing before it goes out, every time. If you want a system that rewrites your site unattended, this is not it, and we would rather say so than quietly become it.
The traffic half needs none of that. AXO installs as one line of script on any site, takes about 15 minutes to go from tag live to first result, and requires no PII. If Shopify's categories look like yours, the cheapest first move is to stop counting AI arrivals as organic and find out whether they behave differently for you too.