Ecommerce AI bot testing checks whether search and answer systems can reach your store, extract accurate product facts, connect those facts to buyer questions, and cite or recommend the right page. It is not a single prompt in ChatGPT. Our ecommerce SEO services use the same evidence-first sequence across crawling, structured data, content and measurement.
Quick answer: in 45 minutes, test one product, one category and one advice page through five gates: HTTP access, useful HTML, structured product data, answer quality and citation behavior. Record the evidence in a scorecard. Fix the earliest failed gate first, because later prompt tests are misleading when a bot cannot retrieve or understand the source.
What an AI bot audit can and cannot prove
A controlled audit can prove that a public URL returns a particular response, contains specific facts, exposes valid markup and appears or does not appear in a dated set of answers. It cannot prove a permanent “AI ranking.” Generated answers vary by model, location, freshness, prompt wording and available sources.
Use the audit to make technical and editorial decisions. Do not turn ten prompts into a universal visibility percentage without explaining the sample.
Prepare three representative URLs
Choose pages that matter commercially and expose different templates:
- Product: a stocked, priority item with variants and reviews.
- Category: a collection that targets a clear product class.
- Advice: a buying guide or comparison connected to those products.
Avoid testing only the homepage. It can pass while every product request is blocked, canonicalized incorrectly or rendered without useful detail.
Minute 0 to 8: test crawler access
Open robots.txt and identify rules for the systems you care about. OpenAI’s bot documentation separates OAI-SearchBot, ChatGPT-User and GPTBot by purpose. Perplexity also publishes crawler and user-agent guidance. Test the right agent instead of a generic label such as “AI bot.”
For each URL, record:
- Initial and final HTTP status
- Redirect path
- Response body size
- Whether the body is the real page or a challenge
- Canonical URL in the returned HTML
- Robots meta directives
If the request is denied, inspect CDN, firewall and server logs. A browser returning 200 while a bot receives 403 is an access failure, not a content problem.
Minute 8 to 15: inspect what the response actually contains
Search the raw response for the product name, description, price context, availability, brand, key specifications, category name and main article answer. Then compare it with the rendered page.
Flag any fact that appears only after a click, geolocation choice, app request or client-side interaction. The goal is not to remove useful interactivity. It is to ensure the public source delivers enough stable information to identify and evaluate the resource.
If the raw response is an empty application shell, document that finding before changing copy. If it contains the full content but the bot is blocked, fix access. If it is accessible and complete, continue.
Minute 15 to 23: compare visible product data and structured data
Run the product page through a structured-data validator and inspect the JSON-LD manually. Google’s Rich Results Test is useful for supported Google markup, while Schema.org vocabulary can help inspect properties beyond a specific rich result.
Compare these fields with what a shopper sees:
| Field | Pass condition | Common failure |
|---|---|---|
| Name | Same product and variant identity | Generic template title |
| URL | Canonical purchasable page | Parameter or wrong market URL |
| Price and currency | Current visible offer | Stale app or feed value |
| Availability | Matches inventory state | Schema says in stock when page says sold out |
| Brand and identifiers | Accurate and consistent | Missing GTIN, MPN or brand |
| Ratings | Eligible visible reviews | Markup from a different product |
Structured data helps machines parse facts. It does not excuse hidden, inaccurate or thin visible content.
Minute 23 to 30: test extractable answers
Write five questions a buyer should be able to answer from the three pages. Good tests are concrete:
- Who is this product best suited to?
- What important limitation should a buyer know?
- Which variant fits a defined need?
- How does it differ from the closest alternative?
- Can it be delivered to the target market?
Read only the page content and answer each question. If a careful human cannot find a supported answer, an AI system has no reliable source either. Mark missing facts, buried facts and contradictory facts separately.
This step often reveals the highest-value content work. A shorter product page with dimensions, compatibility and honest limitations can outperform a long page filled with adjectives.
Minute 30 to 38: run a controlled prompt set
Use fresh sessions and the same prompt wording across the systems you are comparing. Test three intent levels:
- Entity: “What is [brand or product]?”
- Category: “What should I look for in [product type]?”
- Recommendation: “Best [product type] for [specific constraint] available in [market].”
For each response, record the platform, date, model if shown, market, prompt, products named, sources cited, accuracy and whether your page was retrieved. Do not edit the prompt halfway through and compare the result as if it were the same test.
Minute 38 to 45: score the failure layer
| Area | 0 | 1 | 2 |
|---|---|---|---|
| Access | Blocked or challenged | Intermittent or template-specific | Stable 200 response |
| HTML | Core facts absent | Some facts delayed or hidden | Decision facts present |
| Data consistency | Major conflicts | Minor gaps | Page, schema and feed align |
| Answer coverage | Questions unanswered | Partial answers | Clear supported answers |
| Entity trust | Identity unclear | Basic identity only | Consistent business evidence |
| Prompt visibility | Absent or inaccurate | Named without citation or fit | Accurate and relevant appearance |
The maximum is 12. The number is useful only with the evidence beside it. A score of 8 caused by access failure needs a different plan from a score of 8 caused by weak comparison content.
Use a failure taxonomy, not one AI visibility label
Access failure
The bot receives a denial, rate limit, challenge, login or unusable response. Fix robots, hosting, CDN or application delivery. Start with the ChatGPT crawl troubleshooting guide.
Extraction failure
The request succeeds, but essential facts are absent from the delivered content or hidden in images and interaction-only components. Fix template output and accessible text.
Consistency failure
The page, structured data, feed or market variant disagrees. Establish one source of truth, then update every output.
Evidence failure
The product is identifiable but there is no supported reason to choose it for a use case. Add specifications, comparisons, testing, limitations and genuine review detail. Use the diagnostic for stores missing from AI product recommendations to inspect that evidence layer.
Authority failure
The site makes claims that are not corroborated, or the business identity is unclear. Improve policies, authorship, entity consistency and earned third-party evidence.
Demand mismatch
The tested prompt does not fit the product, stock or delivery market. Choose defensible queries instead of forcing relevance.
How to test a large catalogue without checking every page
Sample by template and risk. Include a simple product, variant-heavy product, out-of-stock item, sale item, international page, category with filters, and a guide with product links. Add pages from the highest-revenue and fastest-changing groups.
If one shared template emits the same broken canonical on 4,000 products, the unit of work is the template. If only imported products have missing identifiers, segment by data source. Sampling should reveal systems, not produce a random pile of screenshots.
What to monitor monthly
- Verified AI crawler requests and response codes
- Changes to robots.txt and security rules
- Structured-data validity on sampled templates
- Product feed and visible-page mismatches
- A stable set of buyer prompts and cited sources
- AI referral sessions where analytics identifies them
- Branded search, assisted conversions and revenue quality
Google’s AI search feature guidance says ordinary Search technical requirements remain important. Keep Search Console indexation, clicks and query data in the same review rather than creating an isolated “AI” dashboard.
Turn the audit into an implementation ticket
Each failed check should become a ticket with an affected template, example URLs, current evidence, expected behavior, owner and acceptance test. For example: “Product JSON-LD reports the default variant price after a shopper selects another variant. Update Offer output to match the canonical variant strategy and visible price, then validate five sampled products.”
If your test uncovers a cross-template issue, the ecommerce SEO fix sprint is designed for a confirmed technical or on-page problem that needs implementation and verification. If the failure is still unclear, start with an ecommerce SEO audit.
FAQs
What is ecommerce AI bot testing?
It is a structured check of crawler access, delivered content, product data, answer coverage, citations and recommendation behavior across representative store pages.
Can an AI visibility tool replace server-log testing?
No. A visibility tool can sample answers, but server and CDN logs show whether verified requests reached the site and what response your infrastructure returned.
How many prompts should I track?
Start with 10 to 20 high-value prompts grouped by product, audience, use case and market. Stability and relevance matter more than a large unmaintained list.
Should I test ChatGPT, Perplexity and Google separately?
Yes. Their crawlers, sources, interfaces and answer behavior differ. Use the same core questions but record each system independently.
What is a good AI visibility score?
There is no universal score. Define transparent checks, preserve the evidence and compare your own baseline over time. Tie improvements to accurate discovery and commercial outcomes.
Leave a Reply