In this article
The practical answer: An AI recommendation is only useful if the buyer can take the next step. Audit the whole journey: discovery, accurate comparison, a working destination and a clear path to buy or enquire.
The useful idea behind the viral shopping test
In a LinkedIn post shared with us, Olivia Moore described using a fresh account to watch a shopping agent find a winter coat. She reported that a bot challenge pushed the agent away from Patagonia's own site and toward another retailer, while verifiable availability, pickup and price helped REI win the purchase path.
That is a useful diagnostic story. It is one reported journey, which we have not independently reproduced. It does not establish that those retailers behave the same way for every agent, or that every platform follows the source sequence described in the post.
The business opportunity is concrete: find the points where an otherwise qualified buyer loses access, confidence or momentum. For a retailer, that may be a wrong variant or stale stock. For a service firm, it may be an offer that never explains scope or a booking page that cannot be reached.
Separate three questions before changing anything
Can the system find you? Record whether your brand, product or offer appears, and which page is cited. Can it describe you accurately? Check price, fit, exclusions, stock and policies against current first-party information. Can the buyer proceed? Follow the recommendation to a useful product page, cart, enquiry or booking step.
A mention can pass the first question and fail the other two. Treat those as different problems with different owners; a content rewrite will not repair a broken cart.
Run a small, controlled baseline
Choose one commercially important offer and two AI experiences your customers plausibly use. Access, pricing and available actions vary; this can be a low-cost exercise, but it is not universally free. You do not need an account on every new agent platform.
Write five non-branded prompts from customer questions. Include a use case, a budget, a constraint, a comparison and a location or delivery requirement. Add a separate branded prompt to test factual accuracy. Do not introduce your company into the non-branded test.
Use a clean session and turn off available memory or personalization controls. Record the account state, tool or model, date, region, exact prompt and follow-ups. A new email does not eliminate personalization. Prefer explicit location settings and a stated delivery destination; a VPN alone is not a reliable simulation of a local buyer.
Repeat the same prompts three times per surface. Keep the wording and constraints stable. Save a screenshot or transcript and the visible source URLs for each run. This is a small operational sample, not a market-wide visibility study.
Ask for evidence you can verify
Find three options for [customer job] within [budget] that meet [constraints] in [location]. Link to the pages supporting each option. Distinguish verified facts from estimates. Compare the stated criteria. Stop before submitting an enquiry, booking, payment or order.
After the initial result, ask which pages support a particular claim, which criteria were used in the comparison, and whether any visible page errors prevented verification. If a browsing agent exposes its actions, record them. Not every assistant provides a browser trace.
Its explanation of why it recommended an option is a lead for investigation, not a verified account of internal ranking weights or training sources. Do not publish a guessed algorithm as a fact.
Turn observations into an owned fix list
| Observed problem | Check next | Likely owner |
|---|---|---|
| The offer never appears | Indexability, relevant pages and available catalog coverage | Search or content lead |
| The recommendation gets facts wrong | Conflicting pages, catalog fields and source freshness | Product or content owner |
| A challenge or error blocks the page | Request logs and the verified agent's documented access requirements | Engineering or security |
| The buyer cannot finish | Variants, cart, forms, booking and mobile behavior | Commerce or operations |
Keep fraud protection and authentication in place. Investigate a specific failed request before changing bot rules. OpenAI documents separate controls for its search crawler and training crawler; allowing search access does not require choosing the same setting for training. Read the crawler guidance.
Google says its AI search features retain the usual SEO foundations and do not require special AI markup. Start with useful pages, accessible information and internal links rather than a promised ranking shortcut. Read Google's guidance.
Measure what happened after the mention
Use a row per run: prompt ID, surface, date, brand mentioned, cited URL, factual error, visible block, recommended seller, next-step result and evidence link. Keep blocked and incomplete runs visible.
For each surface, report mentions out of completed eligible tests, plus failures separately. Add the number of inaccurate claims and the proportion of attempted journeys that reached the intended next step. Track qualified enquiries or purchases from identifiable AI referrals where available; untagged and assisted journeys will not all be attributable.
An illustrative result such as 4 mentions in 15 completed runs is a sample observation, not a market share. Compare a later run using the same prompt set. A change in results suggests what to investigate; it does not prove your edit caused it.
Choose one repair before buying another tool
Start with an access failure, a wrong commercial fact or a broken high-intent next step. Then improve the evidence buyers need to compare you: specific use cases, honest trade-offs, verified proof and clear terms. An unranked page can still be a stronger sales asset after these repairs.
Use the free Growth Leak Audit to map the wider journey. If AI discovery is where you see friction, bring your prompt log and evidence to a fit conversation. We can agree whether a focused Growth Fix is appropriate. The scope is agreed first; an AI ranking or revenue gain is not promised.
For store-level implementation, read EcomCoach's product and checkout readiness guide. For a repeatable reporting method, use MarketingCoach's measurement plan.
Sources and editorial note
Sources checked October 20, 2026. The inspiration is a reader-supplied LinkedIn excerpt by Olivia Moore (profile); the original post permalink was not supplied. Her reported shopping results are attributed observations. The audit methods and hypothetical examples here are our editorial recommendations, not measured client results.
Want this applied to your business? The Growth Fix is one fixed-price session to find where your customers drop off and fix the part that matters most.
