Short answer: an e-commerce brand raises its exposure in AI shopping answers by working on four things in order: letting the assistants' search crawlers read its product pages, getting structured product data into the assistants' shopping systems, writing product facts as sentences an assistant can lift, and making sure the marketplaces and review sites that get cited carry the same facts. Then it measures, per engine, with buyer questions that do not contain the brand name. None of this guarantees a recommendation; it makes the brand eligible to be one.

This is the practical companion to why consumers are increasingly relying on AI for shopping. That piece covers the demand side. This one covers what a store can actually change, with the Shopify and WordPress specifics that tend to trip people up.

Where AI shopping recommendations get their products from

Product recommendations in an assistant draw on two different pipes, and most GEO advice only talks about the first.

  • Crawled pages. Search agents such as OAI-SearchBot and PerplexityBot fetch and index public pages, and the answer can quote or cite them. This is where your product pages, buying guides and third-party reviews compete.
  • Structured product data. The assistants now also take product data directly. OpenAI publishes a product feed specification for ChatGPT, Perplexity runs a free merchant program in which retailers share product data, and Shopify syndicates titles, descriptions, images, pricing, inventory and shipping through Shopify Catalog to ChatGPT, Microsoft Copilot, AI Mode in Google Search and the Gemini app.
Two ways a product reaches an AI shopping answerPipe 1: crawled pagesProduct pages, buying guides,marketplace listings, reviewsRead by search agents:OAI-SearchBot, PerplexityBot, Claude-SearchBotGate: robots.txt, firewall, HTML without JavaScriptPipe 2: structured product dataTitle, description, image, price,availability, brand, sellerSent through:OpenAI feed, Perplexity merchant program, Shopify CatalogGate: completeness and consistency with the live pageThe assistant’s shopping answerQuotes and cites pages; shows product cards from the data
Figure 1. A store has two jobs: be readable as a page and be complete as a record. Most GEO advice covers only the left-hand pipe. Source: OpenAI and Perplexity crawler documentation, the OpenAI product feed specification, Perplexity’s merchant program announcement and Shopify’s agentic commerce guide.

Shopify is unusually direct about what happens next. Its agentic commerce guide says ranking factors vary by AI platform but can include data quality, relevance, availability, pricing and engagement signals, and that the quality and completeness of product data decides whether a product surfaces in a conversation or gets passed over. That gives you two jobs, not one: be readable as a page, and be complete as a record.

It is also worth being clear about scale before you plan the work. Adobe reports that generative AI traffic to retail sites grew 693.4% in the 2025 holiday season, but that the base of users remains modest. Shopify reports that in Q1 2026 AI-driven traffic to Shopify stores grew 8 times year over year and orders from AI-powered searches nearly 13 times, which again says more about the growth rate than the size. Treat this as a channel growing quickly from a small base, not as one that already outsells search.

Three claims to discount before you start

How much less likely AI-referred visits were to convert than other trafficU.S. retail sites, Adobe AnalyticsJuly 202443% less likelyEarly 20259% less likely
Figure 2. The gap narrowed sharply but had not closed by early 2025. Adobe reads it as AI being used at the research and consideration stage. Source: Adobe Analytics blog, March 17, 2025.
  • "AI visitors convert better." Adobe's early 2025 data says they converted 9 percent less often than other traffic, although the gap had narrowed sharply from 43 percent in July 2024. Adobe reads the remaining gap as AI being used at the research and consideration stage, so plan for a research-stage channel.
  • "Schema markup makes you cited." None of the crawler documentation linked in this article says so. Use markup to remove ambiguity and keep it consistent with the page and the feed.
  • "Small brands can leapfrog leaders with a few tactics." Sometimes a niche sub-question is open, but in consolidated product categories the realistic goal is to own specific constraints, not the head term. Step 4 shows how consolidated some categories already are.

Step 1: Let the right crawlers in, and test it

Each assistant separates crawlers by job. OpenAI's bot documentation lists GPTBot for training, OAI-SearchBot for the index behind ChatGPT search, and ChatGPT-User for fetches a person asked for. Perplexity's crawler guide separates PerplexityBot from Perplexity-User in the same way. Blocking a training agent is a defensible choice that does not remove you from today's answers. Blocking a search agent does.

AgentJobWhat blocking it costs
OAI-SearchBotChatGPT search indexRemoval from ChatGPT search results
PerplexityBotPerplexity answer indexRemoval from Perplexity answers, visible in citation data almost immediately
Claude-SearchBotClaude search resultsYour pages stop informing Claude's search results
Google-ExtendedUsage control, not a crawlerRemoval from grounded Gemini answers; Google Search is unaffected
GPTBot, ClaudeBotTrainingNothing in today's answers

On Shopify, robots.txt is generated for you and is not permissive by default: it disallows /admin, /cart, /checkout, /search, /policies/ and filtered collection URLs. To change it, add a robots.txt.liquid template to the published theme that outputs Shopify's default groups and then appends your own rules, as described in Shopify's robots.txt.liquid documentation. Do not replace the file wholesale, because Shopify updates the defaults:

{% for group in robots.default_groups %}
  {{- group.user_agent -}}

  {% for rule in group.rules %}
    {{- rule -}}
  {% endfor %}

  {%- if group.sitemap != blank -%}
    {{ group.sitemap }}
  {%- endif -%}
{% endfor %}

User-agent: GPTBot
Disallow: /

Only the published theme is served, so a template in a draft theme does nothing. On WordPress, the virtual robots.txt only disallows /wp-admin/ by default, but check Settings, Reading: the "discourage search engines" box also disables core XML sitemaps, according to the WordPress 5.5 sitemap announcement, and a staging site promoted with that box ticked loses its sitemap silently.

Then test with the real user agent instead of reading the file. A CDN or firewall rule that returns 403 to these agents is a block in practice, whatever robots.txt says:

# Compare a browser fetch with the search agents that build answer indexes
for ua in "Mozilla" "OAI-SearchBot" "PerplexityBot" "Claude-SearchBot"; do
  printf "%s: " "$ua"
  curl -s -o /dev/null -w "%{http_code}\n" -A "$ua" https://yourstore.com/products/example
done

# Then check that the facts are in the HTML itself, without JavaScript
curl -s -A "OAI-SearchBot" https://yourstore.com/products/example | grep -i "return

Our AI crawler access guide lists every documented agent and its source.

Step 2: Get complete product data into the assistants' systems

This is the step that did not exist a year or two ago, and it is the most mechanical.

  • Shopify merchants. Shopify says Agentic Storefronts is automatically activated for eligible merchants and is managed under Settings, Sales Channels, Agentic Storefronts. Open it and confirm your store is eligible and connected. Purchasing through these channels is currently available to U.S. buyers only.
  • Merchants on other platforms. OpenAI's feed requires nine basic fields: item_id, title, description, url, brand, seller_name, image_url, availability and price. It asks for titles of at most 150 characters and descriptions of at most 5,000, and says to keep price and availability current. Search eligibility is on by default, but eligibility does not guarantee display.
  • Perplexity. The merchant program is free and is built on retailers sharing product data.

Completeness matters more than cleverness here. Use the real brand name and seller name, write descriptions as plain factual text, and make sure the price and stock status in the feed match the live product page. A feed that says in stock while the page says sold out is exactly the kind of contradiction a data-quality ranking factor will punish.

Product pageProduct feed or catalog
Who reads itSearch agents crawling your HTMLThe assistant's shopping system, directly
What it must containThe facts buyers filter on, written as sentences and present without JavaScriptEvery required field: for OpenAI, the nine basic fields; on Shopify, the data Shopify Catalog syncs
Typical failureFacts that live in app-injected blocks, tabs or imagesPrice or stock that disagrees with the live page
Where you control itTheme, robots.txt, CDN and firewall rulesFeed settings, or Settings, Sales Channels, Agentic Storefronts on Shopify

Structured data on the page itself is the same discipline in a different place. Shopify does not inject JSON-LD for you; its structured_data Liquid filter has to be called explicitly in the product template, and it emits ProductGroup rather than Product for any product with multiple variants. An audit that insists on "@type: Product" will report a false failure on every variant-bearing product. WordPress core emits no schema.org markup at all, so it comes from the theme or plugins, and duplicated Organization or Product graphs from two plugins are a more common failure than missing ones. Our Shopify and WordPress platform pages cover each setup in detail.

Step 3: Write product facts as sentences an assistant can lift

Assistants lift sentences that answer a constraint: a dimension, a material, a return window, a compatibility. Category copy written for brand tone gives them nothing to lift, however good it reads. Three habits help:

  1. Put the facts buyers filter on into sentences, not only into a spec table. "The pack fits a full-size laptop and the hip belt detaches for carry-on use" can be quoted. A row in a size chart saved as an image cannot.
  2. Answer the constraint questions on the page. Returns, compatibility, sizing, care, who the product is not for. Write each answer so it makes sense if it is the only sentence quoted.
  3. Check that the facts are in the HTML. Shopify storefront pages arrive as complete HTML, but content injected client-side by apps, such as reviews and Q&A blocks, and anything loaded through the Ajax Section Rendering API, is not in that HTML. On WordPress the risks are plugin-shaped: load-more listings, builder tabs and accordions. If your best answers live in those places, a crawler that does not run JavaScript never sees them.

Writing a buyer guide or comparison page helps for the same reason: it matches the shape of a buying question. Keep it honest about where an alternative suits some buyers better; a comparison that only ever recommends you reads as a sales page.

Step 4: Pick the sub-questions you can win, and fix the pages that get cited

Product categories consolidate in AI answers. In our own Q3 2026 benchmark, the three categories with a clear default answer were all product categories: project management software at 38% of answers, skincare at 35% and electric bikes at 28%, while every local-service category had a leader below 13%.

A buying questionin your categoryDo the assistants nameany brand at all?NoOPEN GROUNDWrite the page that answers itYesIs one brand named in aquarter or more of answers?YesINCUMBENTAnswer a sub-question itanswers badly: a constraint,a budget, a use caseNoCONTESTEDFix the pages that get citedBOTH BRANCHES WHERE A BRAND IS NAMEDList the marketplaces and review sites cited for the question; check they carry your current facts, prices, stock.The quarter-of-answers threshold is the incumbent test used in the text above.
Figure 3. Choosing which buying questions to work on. A leader already named in a quarter or more of answers is an incumbent you displace on a sub-question, not a gap you fill. Source: the steps in this guide.

When one brand already appears in a quarter or more of answers, being merely present is not enough. The cheaper path is to become the answer to a specific sub-question the incumbent answers badly: a constraint, a budget, a use case.

The same benchmark found that in product categories a large share of citations went to marketplaces, review sites and publisher round-ups rather than brand sites. Your product page can be perfect and still lose the citation to a listing page that carries the same facts with third-party framing. So:

  • Find the questions in your category where the assistants currently name nobody. Those are open ground.
  • List the marketplaces and review sites that get cited for your category, and check whether they carry your current facts, prices and availability. Correct the ones you can.
  • Keep one consistent product and brand name everywhere. Spelling variants split your own share across several rows.

Step 5: Measure AI shopping recommendation exposure without flattering yourself

AI shopping exposure is invisible in standard analytics until a click happens, so you have to ask the assistants directly. Four rules keep the numbers honest:

  1. Exclude questions that contain your brand name. In one of our own runs, 13 of 15 generated questions contained the brand name, and the score read 13% when the honest value was zero. Write questions the way a buyer who has never heard of you would ask them, with a situation in them.
  2. Report each engine separately. In our benchmark one model named 3.5 providers per answer and another 2.1. An average hides where the work is.
  3. Count only sources marked in the answer as citations. Counting everything the engine retrieved inflated our own citation figures by two to four times before we fixed it.
  4. Treat a change as real only after it repeats. Answers vary between identical runs. Alert on a change that appears in two consecutive runs, not one.

Keep the question set fixed between runs, and keep product questions separate from brand questions: being named for "best running shoes" and being named for a specific shoe model are different problems. If you want this measured continuously, the InsightWonder e-commerce solution tracks product and category questions per engine with these rules.

Sources

An earlier version of this article carried headline statistics, charts and three brand case studies taken from statistics round-up sites rather than original research. We could not verify them at a primary source, so they were removed; every figure that remains links to the publisher's own page or to our own measurement.