AI answers are already shaping what shoppers consider “good options,” often before they ever see a search results page. If your store isn’t showing up, or shows up with the wrong details, you don’t just lose traffic. You lose the shortlist. This audit is a practical way to see how AI systems describe you today, where they get confused, and what to fix so you’re easier to recommend (and easier to buy from) next month.

What “AI discovery” means for eCommerce (and what to measure)

“AI discovery” is how shoppers find (and decide to trust) your store through AI answer engines like ChatGPT, Gemini, Perplexity, and Google’s AI Overviews. Instead of ten blue links, these systems often return a synthesized answer, a shortlist of recommended brands, a few product picks, a “best for” table, and sometimes citations. For eCommerce teams, the question shifts from “Do we rank?” to “Do we get recommended, described correctly, and sent qualified clicks?”

AI discovery isn’t one surface. The same intent can show up as a conversational question (“what’s the best carry-on for international flights?”), a comparison (“Away vs. Monos”), a constraints query (“carry-on under 7 lbs with laptop sleeve”), or a policy check (“return policy for X brand”). You can “win” one and still be invisible in the others, so an audit needs multiple query types, not just brand-name prompts.

Measure signals that show whether the model can (a) find you, (b) understand you, and (c) justify recommending you:

• Brand mentions and product mentions
• Correct attribute statements (price range, materials, compatibility, sizing, etc.)
• Citations/links (when provided)
• Whether you appear in comparison sets
• Whether the answer is actionable: does it point to a relevant product page, collection page, or policy page?

Treat “representation quality” as a core metric. If an answer engine recommends you but misstates shipping thresholds, return windows, warranty, or product specs, you can get high-intent traffic that doesn’t convert, and support ends up cleaning up confusion. Visibility is only half the job; accuracy is what makes that visibility pay off.

Set up a repeatable audit: prompts → score → fixes → re-test

Run your AI discovery audit like an experiment. Define a standard prompt set, run it on a schedule, score outputs with clear thresholds, make fixes tied to outcomes, then re-test. That creates a baseline, a change log, and a way to see what moved the needle.

Avoid changing everything at once. A monthly rhythm helps isolate what helped, updating a policy page, improving product attribute coverage, tightening internal linking, and keeps work distributed across teams (SEO, merchandising, lifecycle) using the same rubric.

You need three assets:

1. A prompt library (your test harness)
2. A scoring sheet (your measurement layer)
3. A fix tracker (owner + hypothesis + re-test date)

Your standard prompt set (brand, category, comparison, “best for”, policy)

Your prompt set should mirror how shoppers ask. Keep it small enough to run monthly, but broad enough to cover revenue-driving intents. For most stores, 25–40 prompts is a workable starting point. Consistency matters. Use the same prompts every month so changes in results are interpretable.

Use five buckets, written in plain language:

1. Brand prompts (navigational + trust)

• “Is [Brand] legit?”
• “Where can I buy [Brand] [product type]?”
• “What is [Brand] known for?”
• “Does [Brand] have a warranty?”
• “How long does [Brand] shipping take?”

These test whether the model can identify your store, summarize your value proposition accurately, and point to the right official pages for policies and support.

2. Category prompts (non-branded discovery)

• “Best [category] for [use case]”
• “Best [category] under $[price]”
• “Most durable [category]”
• “Best [category] for beginners”
• “What should I look for in a [category]?”

These show whether you’re present when the shopper hasn’t chosen a brand yet, and which attributes the model thinks matter.

3. Comparison prompts (shortlist moments)

• “[Your brand] vs [Competitor]”
• “Alternatives to [Competitor]”
• “Which is better: [Product A] or [Product B]?”
• “Top brands like [Your brand]”

These reveal whether you appear in the same consideration set and whether the model can make accurate, differentiating statements.

4. “Best for” prompts (constraint-based buying)

• “Best [category] for [specific constraint]”
• “Best [category] for [body type / size / skin type]” (where relevant)
• “Best [category] for [compatibility requirement]”
• “Best [category] for [climate / environment]”

These matter because answer engines often respond with structured lists. If your product data is incomplete, you can get filtered out before you’re considered.

5. Policy prompts (conversion blockers)

• “Return policy for [Brand]”
• “Does [Brand] offer free returns?”
• “How to exchange [Brand] [product]”
• “How to contact [Brand] support”
• “Does [Brand] ship internationally?”

Policy prompts are where representation errors hurt most, and they’re often the fastest to fix with a single canonical page.

Two rules make this more useful:

• Include at least one prompt for a top collection/category page and one for a hero SKU, because AI can treat “brand,” “category,” and “product” as different entities.
• Write prompts the way customers speak, including messy phrasing.

How to control variables so results are comparable month to month

AI outputs vary. You can’t remove variability, but you can control enough to make results usable.

Standardize what you can:

• Environment: same geography (or record location), same language, same account state (logged out vs logged in).
• Procedure: same tools each time (e.g., ChatGPT, Gemini, Perplexity, and AI Overviews if you can access it), same prompt order, same capture method.

Save screenshots or export transcripts with timestamps. If the engine provides citations, copy the URLs exactly. If it doesn’t, record “no citations provided,” not “no sources.”

Lock your prompts. Small wording changes can produce different lists and citations. If you do change prompts, version the set and treat it as a new baseline.

Do the same for products. Pick a stable set of SKUs to track (one hero, one mid-tail, one variant-heavy product) and don’t swap them monthly unless the assortment changes materially.

Separate model variance from site changes by repeating a subset. Re-run 5 core prompts twice in the same session and note how much the answer shifts. If outputs swing a lot, lean on aggregate scoring across many prompts, not any single “hero” prompt.

Score what AI returns: a rubric with pass/fail thresholds

A useful audit needs a rubric that turns messy language outputs into consistent scores. You’re not grading the model. What you’re grading is your store’s discoverability and representation inside those outputs.

Use pass/fail thresholds, with optional quality points. Pass/fail keeps prioritization clean because it maps directly to actions. Score per prompt, then roll up by bucket (brand/category/comparison/best-for/policy) and by engine (ChatGPT/Gemini/Perplexity/AI Overviews). Keep it simple enough to score in under a minute per prompt.

1. Mention (Pass/Fail)

• Pass: Your brand/store is mentioned when it’s relevant to the prompt.
• Fail: You’re absent from a shortlist where you should plausibly appear, or the model confuses you with another entity.

2. Correctness (Pass/Fail)

• Pass: Key facts stated about you are correct (what you sell, broad price positioning, key product attributes, availability, policies).
• Fail: Any major factual error (wrong return window, wrong materials, wrong compatibility, wrong product line).

3. Citations/links (Pass/Fail)

• Pass: The answer includes a link or citation to an official page (your domain) when it references your store, product, or policy.
• Fail: No link when competitors get links, or links point to third-party summaries instead of your canonical pages.

4. Shoppability / next step (Pass/Fail)

• Pass: The answer gives a clear path to a product page, collection page, store locator, or policy page that matches the intent.
• Fail: The shopper is left with generic advice or is sent to irrelevant pages.

5. Competitive framing (Pass/Fail)

• Pass: When compared, your differentiators are stated in a way that’s accurate and meaningful (not generic).
• Fail: The model assigns you a weak or incorrect position (“budget” when you’re premium, “slow shipping” when you’re not, etc.).

After pass/fail, add two optional quality metrics:

• Attribute coverage (0–2): Does the answer mention the attributes that drive conversion in your category (sizes, materials, compatibility, warranty, certifications, care instructions, etc.)?
• Entity clarity (0–2): Does the model clearly distinguish your brand, your hero product line, and your store domain, or does it blur them with resellers, marketplaces, or similarly named brands?

The output you want: pass rate by bucket, pass rate by engine, and the top failing prompts with the highest commercial value.

Worked example: why AI recommends competitors (and how to fix it)

Imagine a Shopify-style store selling premium water bottles. You pick one hero product (a 32oz insulated bottle with multiple lid options) and one category (insulated bottles). Your prompt set includes: “best insulated water bottle for travel,” “insulated bottle that fits in cup holder,” “[Brand] vs Hydro Flask,” and “return policy for [Brand].” You run these across multiple answer engines and capture outputs and citations.

In the first run, competitors are recommended consistently, and your store is missing, or mentioned without a link. On the scoring sheet, you fail:

• Mention for non-branded category prompts
• Citations even when you are mentioned
• Attribute coverage because answers focus on lid leakproof-ness and cup-holder fit but don’t connect those attributes to your products
• Policy accuracy because one engine invents a return window that doesn’t match your site

Start with the simplest diagnosis. The model can’t confidently extract and verify what it needs from your canonical pages.

For category prompts, answer engines often assemble lists from sources that are easy to parse: comparison articles, retailer roundups, and pages with clear specs. If your PDPs bury key attributes in images, use inconsistent naming for variants, or don’t present compatibility constraints in text, you’re harder to include in a structured list. If your return policy is spread across multiple pages or hard to summarize, the model may guess.

Map each failure to a fix that improves extractability and reduces ambiguity:

• Missing from “best for” lists: Add a short, plain-language “specs at a glance” section on the PDP and category page that includes the attributes those prompts care about (dimensions, weight, cup-holder fit, insulation time if you state it, lid types, leakproof claim if you can support it). Make sure the text is in HTML, not only in an image.
• No citations to your domain: Create or strengthen canonical pages that are clearly the “source of truth” (a single returns page, a single warranty page, a single shipping page), and link to them consistently from header/footer and relevant PDP sections.
• Comparison prompts favor competitors: Publish a neutral comparison page that helps shoppers choose based on constraints (not a takedown), and link it from your category hub. Even if the model doesn’t cite it directly, you’re giving it a clean narrative it can reuse.
• Policy inaccuracies: Rewrite policy pages in scannable sections with unambiguous statements (time window, condition requirements, who pays return shipping, exceptions). If there are exceptions, list them clearly.

Teams often focus on schema and forget the on-site experience that turns AI-referred traffic into buyers. If shoppers arrive with context from an AI answer, your pages need to confirm key points fast: what’s different, which option to pick, and what happens if it doesn’t work out. The format can vary; the goal is content that’s explicit, consistent, and easy to restate.

When you re-run the same prompts next month, look for movement tied to your hypotheses: higher mention rate in category prompts, more citations to your canonical pages, fewer policy errors, and more accurate attribute-based recommendations. If those don’t change, revisit the diagnosis. Entity confusion (brand name overlap), reseller dominance, or missing merchant feed attributes rather than on-page content.

Beyond schema: improve structured extractability and entity clarity

Schema helps, but schema alone rarely fixes AI discovery for stores. Answer engines need confidence, and confidence comes from repeated, consistent signals. The same product names, the same attribute statements, the same policy language, and clear canonical pages that can be cited.

Think in terms of structured extractability. How reliably a system can pull a fact from your site (what the product is, who it’s for, what variants exist, what constraints apply, and what happens after purchase). Improve it with clear page structure, consistent headings, scannable specs, and by avoiding critical information that exists only in images, PDFs, or widgets that don’t render as text.

Entity clarity is the other half. AI systems build an internal map of entities, your brand, your store domain, your product lines, your hero SKUs, and sometimes founders or retail locations. Mixed signals lead to mixed results. Common sources of confusion include inconsistent naming between PDPs and collections, multiple “official” policy pages, product lines that share names with generic terms, and third-party marketplaces outranking your domain for brand queries.

Two practical improvements that usually help both:

• Create canonical “source of truth” pages and link to them everywhere. One shipping page, one returns page, one warranty page, one contact page. Make them easy to find from every PDP and from the footer. If you have region-specific policies, make the differences explicit and keep the structure consistent.
• Standardize product naming and attribute language across templates. If you call it “32oz” on one page and “1L” on another, or “stainless steel” vs “steel,” you’re adding ambiguity. Pick a standard and apply it across PDPs, collection filters, and feed attributes.

Also watch for “thin” pages that AI might cite because they’re easy to parse, even if they’re not your best pages. A barebones FAQ can get cited more than a richly designed PDP because it’s straightforward, so make sure your best pages are also clear and extractable.

If you want to connect this audit to a broader AI plan across acquisition and conversion, the sibling post “The AI Stack for eCommerce Marketers” is a useful map of where AI touches the funnel and what teams typically own each layer. Take a look at it here. 

The merchant-surface audit: feeds, variants, and policy data AI can shop

Many AI discovery outcomes don’t come from your blog content or category pages. They come from “merchant surfaces”. Product feeds, structured catalogs, and shopping integrations that power product-level recommendations. If your feed is incomplete, messy, or inconsistent with your site, you can disappear when the engine is trying to return a shoppable answer.

Audit your product data the way a shopping agent would. Pick 20 SKUs: top sellers, a few long-tail products, and at least five variant-heavy items (size/color/material). For each, check whether core attributes are present, consistent, and machine-readable across three places:

• the PDP
• the collection listing
• the feed/export your team uses for shopping channels

Look for gaps like missing GTINs, unclear variant naming, inconsistent sizing, missing materials, missing compatibility constraints, and missing images per variant.

Variants are where many stores lose the plot. If a product has multiple sizes or configurations, the model needs to understand what changes and what stays the same. If variant titles are “Default Title,” if size is only shown in a dropdown without descriptive text, or if key constraints are only in an image, you make it harder for systems to recommend the right version. That leads to two outcomes. You don’t get recommended, or you get recommended but the shopper lands on the wrong variant and bounces.

Policy data is part of the merchant surface too. Shoppers ask “free returns?” and “how long is shipping?” because those are conversion gates. If policy pages are unclear, inconsistent by region, or not linked from product pages, the model may omit you from recommendations (“unclear returns”) or misstate your terms. Treat policy pages like product data: canonical, consistent, and easy to cite.

Run this audit by scoring attribute completeness per SKU and per category, then rolling it up into coverage rates you can improve over time. You don’t need fancy tooling to start. A spreadsheet with columns for category-relevant attributes and a present/missing flag will quickly show patterns like “materials missing on 60% of SKUs” or “compatibility only described in images”, the kind of work that tends to improve AI outputs.

Prioritize fixes like a revenue team: impact/effort tied to outcomes

After your first audit run, you’ll have a long list of failures. The difference between a useful audit and a forgotten spreadsheet is prioritization. Score each fix by expected impact on outcomes (mentions, citations, qualified AI referrals, conversion rate from AI traffic) and by effort (time, dependencies, risk).

Tie every fix to a measurable outcome you can re-test next month:

• “Rewrite returns page” → expected outcome: higher policy accuracy pass rate; more citations to your domain on policy prompts
• “Add specs-at-a-glance to PDP template” → expected outcome: improved attribute coverage scores; increased inclusion in “best for” lists
• “Standardize variant naming and attributes” → expected outcome: fewer incorrect variant recommendations; better shoppability/next-step scores
• “Create comparison page for top competitor” → expected outcome: improved competitive framing and citations in comparison prompts

Use a simple impact/effort matrix with a third dimension: breadth (how many prompts and SKUs a fix affects). Template-level changes (PDP spec section, canonical policy page structure, internal linking patterns) usually have high breadth. One-off blog posts usually have lower breadth unless they target a high-value prompt cluster.

A practical scoring model:

• Impact (1–5): How much this could move mentions/citations/qualified traffic
• Effort (1–5): Engineering + content + approvals
• Breadth (1–5): Number of SKUs/pages/prompts affected
• Confidence (1–5): How sure you are the fix addresses the root cause

You don’t need perfect math, just a shared language for tradeoffs. A “high impact, low effort, high breadth” fix should go first. A “high effort, low confidence” fix is a candidate for a smaller experiment (update one category hub first, then roll out).

Keep prioritization honest by separating “visibility work” from “conversion work,” even though they overlap. Some fixes mainly increase the chance you’re cited (canonical policy pages, clearer entity signals). Others mainly improve what happens after the click (clearer PDPs, better variant selection). Both matter; the point is knowing what outcome you’re buying with each sprint.

Make it a monthly cadence: tracking, reporting, and what “good” looks like

A one-time audit creates noise. A monthly cadence creates momentum. The goal is to make AI discovery measurable enough that it becomes part of normal acquisition and conversion reporting, not a side project that only resurfaces when traffic dips.

Start with a lightweight monthly report that includes:

• Pass rate by bucket (brand/category/comparison/best-for/policy)
• Pass rate by engine (so you can spot where you’re consistently weak)
• Top failing prompts by commercial value (tied to your highest-margin categories or hero SKUs)
• A short change log (what you shipped since last run: policy rewrite, PDP template update, feed cleanup)
• A “next month” hypothesis list (3–8 fixes, each tied to a metric you expect to move)

Define what “good” looks like so you’re not chasing perfection:

• Brand and policy prompts: high correctness, high citation rate to your domain, clean next steps
• Category and best-for prompts: improving mention rate over time, plus better attribute coverage (you’re included for the right reasons)
• Comparison prompts: fewer generic statements, more accurate differentiators, and fewer cases where resellers or marketplaces replace your official site

Finally, track outcomes beyond the audit. If you can, tag and segment AI-referred traffic in analytics so you can watch bounce rate, PDP engagement, and conversion rate for those sessions. The audit tells you what AI says; your site data tells you what shoppers do with it.

A monthly AI discovery audit won’t make every model recommend you overnight. What it does give you is control. A repeatable way to spot where you’re missing from the shortlist, where you’re being described incorrectly, and which fixes are most likely to improve both visibility and conversion over time.