Where AI Agents Remember You, and Where They Search Again
When an AI agent needs a product or a service, does it go back to a source it knows, or search from scratch? It does both. An agent works in five layers, and only one of them starts over on every question.
| Layer | What it does | Sticky? |
|---|---|---|
| Web retrieval | Searches the web and cites sources for this answer | No |
| Model knowledge | Knows which brands belong in an answer from its training data | Yes, until the next model |
| User memory | Remembers one person's preferences across chats | Yes, for that person |
| Connected tools | Calls an app or service the user has connected | Yes, while it stays connected |
| Platform feeds | Reads product data a platform sends it | Yes, while you stay eligible |
Web retrieval starts over
SparkToro and Gumshoe had 600 volunteers run 12 prompts through ChatGPT, Claude and Google's AI a combined 2,961 times. They published the results on January 27, 2026. The same list of brands came back in fewer than 1 in 100 responses. The same list in the same order came back in fewer than 1 in 1,000.
The sources don't match Google either. In an Ahrefs study of 15,000 prompts from August 11, 2025, only 12% of the links that ChatGPT, Gemini and Copilot cited ranked in Google's top 10 for the same prompt. About 80% didn't rank anywhere in Google for that query.
So a citation today doesn't hold a position the way a ranking does. The next run draws a new set.
Model knowledge changes slowly
The same SparkToro study found a steadier pattern underneath the churn. Across 994 responses about headphones, with the prompt worded differently each time, Bose, Sony, Sennheiser and Apple showed up in 55% to 77% of answers. The exact list changed on almost every run. The brands the model reached for didn't.
Ahrefs studied 75,000 brands on May 26, 2025, and found what tracks that presence. Brand mentions across the web correlated with AI Overview mentions at 0.664. Backlinks correlated at 0.218. Brands in the top quarter for web mentions averaged 169 AI Overview mentions, against 14 for the next quarter down. That's a correlation, and Ahrefs doesn't claim mentions cause it.
This layer updates when a new model ships with newer training data. A brand that's widely written about before that cutoff goes in with it.
User memory remembers one person
OpenAI's memory FAQ says memory "can help you pick up where you left off and spend less time repeating yourself." Its shopping research page says that with memory on, shopping research "can use what it already knows about you in ChatGPT to better tailor suggestions."
Anthropic brought the same memory to chat and Cowork on August 25, 2026: "wherever you work with Claude, it starts from what it already knows about you."
The vendors document preferences carried forward: interests, sizes, budgets. We found no primary source showing an agent choosing the same store again for a repeat purchase. That part is still an open question.
Connected tools skip the search
OpenAI launched apps in ChatGPT on October 6, 2025. In OpenAI's words, "You can discover them when ChatGPT suggests one at the right time, or by calling them by name." An app the user calls by name doesn't compete in a search result.
Commerce is building the same kind of link. Shopify's Shop Pay handler for its Universal Commerce Protocol lets an agent platform acquire a token for a buyer's Shop Wallet. The platform can use that one Shop Pay integration across all participating merchants.
Platform feeds are pushed in
Some agents don't search for products at all. They read a feed. Shopify sends its merchants' products to ChatGPT, Gemini, Copilot and Meta through Shopify Catalog, on by default for eligible stores. We covered the eligibility rules in Shopify's Four AI Channels Share One Checklist.
OpenAI's own commerce feed spec asks merchants for an initial sample feed, then daily snapshots. It says recommended attributes, "like rich media, reviews, and performance signals," improve ranking, relevance and user trust.
What becoming known looks like
Here's our read, and it's a theory, not a finding. A company that gets cited early hasn't earned much, because citations reset. A company can become the one agents go back to in three ways:
- Get written about widely, so the next model learns the name along with the category.
- Become a tool people connect, so the agent calls you by name.
- Keep your feeds complete, so every platform that reads them can offer you.
All three take longer than a citation, and all three last longer. An early start helps because each one builds on itself: mentions published before the next model's cutoff, and users who already connected your tool.
Where to start
- Count where your brand is mentioned off your own site, not just where you're linked.
- Check whether your products are in Shopify Catalog, Merchant Center, or the feed your platform offers.
- Find the tasks a customer would hand an agent, and ask whether one of them could be an app, an MCP server, or a checkout integration.
- Treat any single AI citation as one sample, and track how often you appear across many prompts.
Become the source agents go back to.
Pixelcove does design, technical SEO, GEO, and Shopify, run on systems we build ourselves. Most engagements start with our $1,500 SEO audit. If you want agents to know your name, let's talk.