The new search landscape
Shoppers ask ChatGPT or Google a question and get three products with reasons attached.
If yours is one of the three, you get a buyer. If it is one of the other two hundred candidates, the shopper never learns you exist.
Any system that produces that answer: ChatGPT, Perplexity, Copilot, Google's AI results.
Each runs its own crawlers, index, and rules - so "we rank well on Google" no longer settles the visibility question.
The AI answer at the top of normal Google results.
It sits above your hard-won first position and answers the question before anyone scrolls down to you.
- Traffic falling while rankings hold? This is usually where it went.
Google's chat search, where the shopper asks follow-up questions instead of new searches.
The comparison, the doubts, and the shortlist can now all happen inside Google before your site gets its first visit.
Four names for one discipline: getting AI systems to recommend you.
Vendors coin new names to look first; the work underneath is the same. Pick one name internally and hold everyone to it.
A search that ends without a click - roughly six in ten Google searches.
The practical shift: you measure whether answers name you, because the visit you used to count often never happens.
How AI finds your catalog
The bots that fetch your site for AI systems: GPTBot, ClaudeBot, PerplexityBot.
Plenty of stores still block the exact bots they want recommendations from.
- First question for your engineers: are we letting them in?
Whether page content sits in the HTML itself or gets assembled by JavaScript in the browser.
Most AI crawlers skip JavaScript, so a beautiful product page can reach them with no price and no product on it.
- One question settles it: what does our page contain with JavaScript turned off?
A model has two memories: training data, frozen months before release, and a live search index.
When ChatGPT quotes the price you changed last spring, the frozen memory won. The fix is making current pages consistent and loud enough for the live memory to win.
The model searching the web at answer time and reading a few pages before responding.
It reads your pages as they are today, on a timeout measured in seconds - so a slow page, or a price behind a click, misses the answer.
The file that tells bots which parts of your site they may fetch.
One wrong line can keep your whole catalog out of AI answers, and the file has usually sat untouched for years.
- Have someone read it this week.
The switch that keeps your content out of Gemini's AI training.
If legal wants you "out of AI," this is the control that does it without touching search visibility. Blocking more than this is how stores hurt themselves.
A proposed file that hands AI systems a map of your site.
Google ignores it, Perplexity reads it, measured impact is modest. Fine to add.
- If a vendor's pitch leads with llms.txt, the pitch has nothing heavier behind it - walk.
Published instructions for AI agents that come to shop: what you sell, how to search your catalog, how to buy.
Adoption is early on both sides. An afternoon of work that answers questions agents otherwise guess at - and no more than that.
The fixed amount of attention crawlers give your site.
Duplicate pages, dead filters, broken links, and stale sitemaps spend it; whatever remains goes to products. This is why technical work comes before content.
The product file you send Google and Microsoft with prices, stock, and variants.
Shopping answers trust the feed over the page - so a feed that silently drops half your variants halves your shelf.
How AI understands your products
Product facts written for machines: price, stock, ratings, in a standard format on the page.
Without it, AI has to guess those facts from your prose - and its guesses become what shoppers hear.
The standard structured-data format for products.
Google validates it and rewards it with rich results: the stars, prices, and stock notes on your listings.
- If the stars disappear, a release broke your product data.
Anything AI tracks as a distinct thing: your brand, a material, a product line.
Recommendations run on connections between entities. If "Supima" and your T-shirts are never linked in machine-readable form, that recommendation goes elsewhere.
The map of entities and how they relate: the AI's picture of your catalog.
- Ask ChatGPT what your brand is known for.
- Ask which of your products suit a need you serve well.
- Wrong answers show you the gap the entity work closes.
One sellable version of a product: size, colour, fit.
The classic failure: the feed exports one variant, the page shows twelve, and AI recommends the sold-out one.
AI expands one shopper question into dozens of small searches about fit, sizing, weather, and stock.
No keyword tool shows these - which is why content planned from keyword lists keeps missing how AI shops.
How AI stores text: as coordinates of meaning, matched by closeness rather than exact words.
A plainly stated fact beats a clever one; the plain version sits closer to how shoppers phrase the question.
AI stating something false about you with full confidence, usually an old price or a dead product.
You cannot stop models from guessing. You can make every source they read tell the same story, so the guess comes out right.
Old SEO terms that still decide AI answers
Live retrieval behaves like a person in a hurry: search, read the top results, answer. The pages that rank are the pages AI reads.
The search engine results page: the list of links after a Google or Bing search.
It anchors this whole section, because live retrieval searches, reads the top results, and answers - ten-year-old plumbing still decides brand-new answers.
Your declaration of which page is the original among near-copies.
Set wrong in one template, it hides a thousand products in a single deploy. Worth an audit after every replatform.
The same content living on many URLs.
Systems pick one winner themselves - and their pick can be your tracking-link version, which is then what AI reads and cites.
Filter pages minting URLs: size 8, blue, on sale, in stock.
One category can breed tens of thousands of near-copies that soak up the crawler attention your products needed.
How crawlers infer importance: pages you link to often must matter.
Check which group your best-margin products are in. The ones three clicks from anywhere lose.
A URL forwarding to a URL forwarding to a URL.
Crawlers give up after a few hops. Chains accumulate through every migration you have ever done, and nobody owns cleaning them up.
The tag that removes a page from search.
Correct on a checkout page; catastrophic when a template applies it to products. Entire catalogs have vanished this way after one deploy.
Links and unlinked mentions from the open web.
Models learn what your brand is from press coverage and expert reviews - on top of their old ranking value, both now shape what AI believes about you.
Google's credibility rubric: experience, expertise, authority, trust.
What you can observe: answer engines cite sources that sound like they have handled the product. Brochure copy does not sound like that.
How AI recommends and sells
The three ways to appear in an answer, in rising value.
Ask any visibility vendor which of the three they count. It is often the cheapest one.
- Mention: named in passing, no link.
- Citation: linked as a source.
- Recommendation: chosen for the shopper - the event this discipline exists to win.
An agent is software that shops on a person's behalf: compares, fills the cart, pays.
It sees your data and your prices, and none of your merchandising. When orders arrive this way, product data, availability, and price do all the selling.
The open standard that lets AI systems use your services directly: search the live catalog, build a cart, check stock.
The reason to want one: assistants answer from your data instead of scraping guesses off your pages. When your engineers say "we need an MCP endpoint," this is what they mean.
Buying inside ChatGPT, built by OpenAI and Stripe: the shopper pays in the conversation, you fulfill the order.
The strategic question it forces: can your products sell well on a page you do not own?
Google's standard for the same job, backed by Shopify, Walmart, and Target.
Two standards now compete for agent checkout. The hedge is unglamorous: product data clean enough to publish to both.
How you measure and produce
Asking AI systems your buyers' questions on a schedule and logging which brands get named.
Answers vary between runs, so treat it like polling: samples and trends.
- Never trust the single screenshot in a vendor deck.
Your percentage of relevant AI answers versus competitors - the successor to rank tracking.
There is no standard methodology yet, so two vendors will bring you two different numbers. Pick one and watch the trend.
Producing content from approved product data through briefs, prompts, and review gates, the way feeds are produced.
When someone proposes "more AI content," this is the counter-question: show me the pipeline, the sources, and the review gate.
What your page says that the rest of the web does not.
AI can restate the consensus without your help, so the only content worth funding is what only you can say: your data, your tests, your customers.
Generic AI-made text with no source data or point of view behind it.
The default output of every "AI content" engagement without a pipeline behind it. Readers and ranking systems are learning to filter it - named by the same internet that named spam, for the same reason.
See which of these terms is costing you revenue.
Book a strategy session and we will walk through where your site stands across visibility, technical setup, content, and ownership.