Particular Audience Particular Audience
Talk to us

The retail operating system

One engine for every retail decision.

Most retailers run four or five separate tools.

Search in one tool, pricing in another, ads bolted on top — nobody owns the order shoppers see, and margin leaks between them. The Unified Decision Engine governs relevance, inventory, margin and sponsored demand in one pass, on every surface, including the AI assistants now doing the shopping.

Start with one surface. Leave the rest of your stack untouched.

Need one thing first? Start with our AI product recommendation engine or AI site search.

Measure Brand Visibility in AI Shopping: A Retail Guide

Build a repeatable AI-shopping visibility study with defined prompts, surfaces, identities and denominators. Separate observed presence from traffic and sales.

How to measure brand visibility in AI shopping

Measure AI-shopping visibility as a defined observation study: a fixed set of shopper tasks, a recorded environment and a consistent rule for identifying products, brands and merchants. The result describes what appeared in that sample. It does not establish a brand’s share of all assistant conversations or the sales caused by appearing.

Define the decision and the observation panel

Begin with a useful decision: which product-data issue to investigate, which merchant offer is missing, or whether a revised catalogue description appears accurately in a defined set of tasks. Choose prompts that represent those tasks before reviewing the results. Keep brand-led queries separate from discovery queries so their different intent is visible.

Record platform, surface, country, language, account context, timestamp and the exact prompt. Where the experience supports it, document whether the run starts without prior conversation context. Do not describe all runs as equivalent when settings or the shopping surface changed. Use manual observation or an authorised collection method with a documented operating limit.

Classify the visible surface

For each valid run, record whether the evidence is an answer mention, a product result, a merchant offer or a labelled advertisement. Preserve a reference or permitted capture sufficient for another reviewer to check the classification. Product identity, brand identity and merchant identity require separate fields; name similarity alone is not a reliable match.

Mark ambiguous matches for review. A failed run, a surface that does not load and a valid run with no relevant result are different outcomes. Keep the first two in the operational failure log and apply the declared denominator rule to the last. This prevents technical failures being silently converted into either successes or product omissions.

Compare like with like

Predefine the task panel and repeat policy before comparing periods. If you change prompts, markets or account settings, identify the break rather than displaying a continuous trend. Record a catalogue change alongside the observation date, but do not claim it caused a visibility change solely because the events occurred in sequence.

Use the sample to identify a next investigation. An incorrect product attribute suggests a product-data check. A missing merchant offer suggests a different diagnosis from an absent product. A clear answer mention without a product link may be useful visibility but is not a measured referral.

Connect observations with merchant evidence

When a shopper reaches the retailer, reconcile available referral, session and order records using declared analytics rules. Keep observed presence, recorded clicks and attributed orders as separate measures. Missing referrer information should remain unknown rather than be assigned to an assistant by assumption.

PA can help frame the retailer’s catalogue and discovery requirements. Its public connector guidance describes sample share-of-voice concepts rather than a verified live external-visibility reporting product. Scope the measurement workflow and ownership before promising a dashboard.

The useful output is an evidence log with actionable issues, not a universal ranking score. Review the highest-impact mismatches with catalogue, analytics and commerce owners, then rerun the same declared panel to assess what changed.

Measure sampled visibility without calling it market share

Choose the unit before calculating a percentage. For run-level presence, count valid runs with at least one qualifying brand product and divide by all valid runs in the declared panel. Count the brand once per run, even when several of its products appear. A valid run with no product result remains a valid run with no brand presence.

For sampled product-result share, divide the brand’s qualifying product entries by all qualifying product entries in those runs. Deduplicate the same product within a run according to the declared identity rule. Keep merchant offers, answer mentions and advertisements out of this product-result denominator.

In a synthetic study, the brand appears in 6 of 10 valid runs: 60% run-level presence. Those runs contain 40 qualifying product entries, of which 12 belong to the brand: 30% sampled product-result share. Neither figure is market-wide share of voice.

Report the panel, failed runs, unresolved identities and counting rules alongside the figures. If there are no valid runs, presence is undefined. If no qualifying products appear, product-result share is undefined. Show those states explicitly rather than displaying a misleading zero.

Separate product selection from merchant-offer visibility

A product appearing in an assistant’s shopping experience does not establish that every seller of it appeared. OpenAI’s shopping guidance distinguishes product information from merchant offers. Record those identities separately so the observation points to the right investigation.

In a synthetic run, an Example Brand mug appears as a product result, but the recorded merchant offers do not include Example Retailer. Mark brand-product presence as observed and Example Retailer offer presence as not observed in that run. Do not turn the latter into a claim that the product was rejected or that the retailer never appears.

An answer mention, a product card, a merchant offer and a labelled ad should have different surface labels. Preserve the evidence needed to check the classification and mark ambiguous cases for review. A mention without a link is not a measured referral; a visible ad is not proof of organic recommendation.

Use the resulting issue type to choose the next check: product identity, merchant offer data, integration status or the observation setup. Repeating a consistent task can clarify the sample without proving a permanent rank or exclusion.

Illustrative example

SYNTHETIC observation-log row uses Example Brand, example.com and a demo capture. It is not a real assistant result. Quantified examples are supplied by the P31 insertion and use declared denominators.

Sources

How Particular Audience does it

Frequently asked questions

Is sampled AI visibility the same as market share?

No. It describes a chosen observation panel and its denominators. A useful report discloses prompts, valid runs, markets, surfaces and matching rules. It should not be extrapolated to all assistant activity without a justified sampling design and relevant population evidence.

Does a missing product in one response prove a feed problem?

No. Treat it as an observation that needs diagnosis. Check product identity, current offer data, integration status and the recorded context, then repeat the defined task. Keep technical rejection evidence separate from a product simply not appearing in a particular result.

Can we attribute sales to a visibility improvement?

A temporal association does not establish causation. Analyse available referrals and orders under a declared attribution model, and use a suitable causal design when the question is additional sales. Keep visibility observations, attributed commerce outcomes and incrementality results in separate fields.

Why can run presence and product-result share differ?

A brand can appear in many runs while occupying relatively few product entries, or appear in fewer runs with several products each time. The units differ. Report both with their numerator and denominator rather than treating one as a substitute for the other.

How should a valid run with no product results be counted?

It remains in the valid-run denominator for presence and contributes no brand presence. It adds no product entries to the product-result denominator. Keep it separate from a technical failure, which is excluded under the declared validity rule and reported in the failure count.

What should the report show when a denominator is zero?

Display not calculable and the reason. No valid runs means run-level presence is undefined; no qualifying product entries means product-result share is undefined. Do not substitute zero percent, because that would imply an observed comparison that did not occur.

What if our brand appears but our retail site does not?

Record product presence and your merchant-offer presence separately. Check the relevant offer’s identity, current data and integration status. The observation is evidence about that run; it does not show that the product failed ingestion or that your store is permanently excluded.

Should a plain-text mention count as a product-card appearance?

No. Keep a separate answer-mention field. A product result, an offer and an answer mention have different visible properties and possible next actions. Combining them without a declared rule makes trends difficult to interpret.

How should a labelled advertisement be recorded?

Give it a paid-placement surface label and keep it outside the organic product-result denominator. Preserve the observed disclosure. If classification is uncertain, flag it for review instead of counting it as both paid and organic visibility.

Related guides

How-to guides · All guides