AI shopping share of voice measures how often your brand or products appear in AI-generated shopping recommendations compared with the competitors you track. To measure it, test a consistent set of commercial prompts across assistants, record product mentions and positions, then calculate your share using a defined counting method. Repeat the tests over time: one answer is a snapshot, not a reliable trend.
This guide walks through a practical process for e-commerce teams, from choosing prompts to interpreting results. If you are setting up AI visibility work for an online store, see SearchFIT's e-commerce resources alongside this measurement framework.
What AI shopping share of voice measures
Traditional search share of voice usually describes how much visibility a brand receives in a search results page or advertising auction. AI shopping share of voice looks at the products and brands an assistant includes in an answer to a shopping-related question. Depending on the interface, that visibility might be a mention in prose, a ranked recommendation, a cited source or a shoppable product card.
The important difference is that AI answers can combine several products into one response. Your brand may appear in a list without being the first recommendation. A product card may be visible even when your brand is not named prominently in the written answer. A useful measurement system records these outcomes separately rather than treating every appearance as identical.
AI shopping share of voice is a comparative metric. It only makes sense when you define the market, competitor set, prompts, platforms and counting rules. If those definitions change between reporting periods, the result may move because your measurement changed, not because the assistants changed their recommendations.
The basic AI share of voice formula
Use this formula for brand or product mentions:
AI shopping share of voice = your tracked brand or product mentions ÷ all tracked brand or product mentions across the competitor set × 100
For example, imagine a test set produces recommendations that include your brand and several competitors. Count each eligible brand mention according to your chosen rule, add the mentions for every brand in the comparison set, then divide your count by that total. The result is a percentage of the tracked recommendation space, not a percentage of all shoppers or all possible AI answers.
There is no single universal counting rule. Some teams count a brand once per answer. Others count every named SKU. Choose one approach, document it and use it consistently. Semrush's guide to measuring AI share of voice also discusses the value of defining a consistent comparison set and measurement method.
Track related metrics separately
Share of voice is useful, but it cannot tell you everything about product recommendation quality. Pair it with other measures so a small increase in mentions does not hide a decline in the quality or relevance of those mentions.
| Metric | What it tells you | What it does not tell you |
|---|---|---|
| Mention rate | How often your brand appears in valid responses | Whether the appearance is prominent or persuasive |
| AI shopping share of voice | Your share of tracked brand mentions relative to the competitor set | Whether the mentions lead to visits or sales |
| Top-pick rate | How often your product is named as the leading recommendation under your rule | Whether the answer is factually accurate or commercially useful |
| Product-card rate | How often a shoppable card for a tracked SKU appears | Whether the card is available to every user or region |
| Citation rate | How often the assistant links to your site or another source you control | Whether the cited page caused the product recommendation |
| Sentiment or framing | Whether the answer describes the product positively, neutrally or negatively | Whether that description reflects the full customer experience |
A mention rate is calculated as responses mentioning your brand divided by valid query runs, multiplied by 100. A top-pick rate uses responses naming your SKU as the leading option divided by valid runs. These are not interchangeable with share of voice: one measures your presence across tests, while the other compares your presence with competitors.
What you need before you start
A spreadsheet is enough to run a small baseline. A tracking platform can help when you need regular tests, multiple brands, larger prompt sets or reports across assistants. Before you run anything, agree on the following with your team.
- A clear category: Define the product range you want to measure. Do not combine unrelated products just to increase the number of prompts.
- A competitor set: Include the brands a shopper might realistically compare with you. Record why each competitor belongs in the set.
- A prompt list: Use prompts that represent genuine shopping needs, not only phrases that contain your brand name.
- A platform list: Decide which assistants and interfaces matter to your customers. Keep platform and mode names specific in your records.
- A counting guide: Define how to count repeated brand mentions, multiple SKUs, cards, citations and products that appear in a comparison table.
- A reporting cadence: Decide when you will repeat the test and who will review results. Regular, comparable checks are more informative than occasional manual searches.
You can use SearchFIT's resources to explore AI search measurement and planning. If you are comparing a manual workflow with a platform-supported one, review SearchFIT pricing and assess the time needed to collect, check and report your own results.
Step-by-step: measure AI shopping share of voice
1. Set the business question and scope
Start with the decision you want the measurement to support. For example, are you trying to learn whether assistants recommend your products for a specific use case? Are you checking if a product feed update changed how your catalogue appears? Or do you need a category-level view for a client report?
Write down the scope before selecting prompts. Include the product category, region, language, audience, competitor set and assistants. If you sell both entry-level and premium products, do not lump them together if shoppers would not compare them directly. Separate those segments or make the distinction clear in your analysis.
Set boundaries for what counts as a valid result. An assistant refusal, a broken response or an answer unrelated to the prompt should not count as a normal recommendation. Record it as invalid or as a separate outcome, and state the reason. Do not quietly remove answers just because they are inconvenient.
2. Build a prompt basket around shopping intent
A prompt basket is a standard set of questions that represents the ways shoppers ask for product help. It gives you something consistent to test each time. Include unbranded category searches, comparisons, constraints and use cases. The mix should reflect your products and customer research, not a generic template alone.
Examples include:
- Unbranded category: Best insulated water bottle for commuting and hiking.
- Comparison: Compare Brand A and Brand B for a lightweight travel stroller.
- Constraint-based need: Find a fragrance-free moisturiser that works for sensitive skin.
- Use case: Recommend a compact espresso machine for a small apartment.
- Budget or feature constraint: Which noise-cancelling headphones have replaceable ear pads?
Adapt each example to your catalogue. Use the language customers use in reviews, support tickets, sales conversations and site search. Include different levels of specificity: a broad category prompt can reveal discovery visibility, while a detailed prompt tests whether an assistant understands product attributes.
Keep most prompts unbranded when you want to measure discovery. If every prompt names your company, you are testing whether the assistant responds to brand demand, not whether it discovers your products for a shopper who has not chosen a brand. Keep branded prompts in a separate group because they answer a different question.
AI Rank Lab's guide to AI share of voice formulas and sample size discusses prompt-set design. Its suggested prompt range can help teams plan a broader basket, but the right size depends on your category, available time and the decisions you need to make. Start with a manageable set that you can repeat and review carefully, then expand when the results are useful.
3. Choose the assistants and test modes
Select platforms your customers are likely to use for product discovery. Your list could include ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews or AI Mode, where available. Do not treat a platform name as a complete test description. Record the interface or mode, account type where relevant, date and region.
These systems do not necessarily select or display products in the same way. One answer may show product cards, another may link to sources, and another may provide a text comparison. Some shopping experiences can use merchant catalogue information or structured product data, while other answers may lean more visibly on web pages and editorial sources. Treat these as differences to observe and record, not assumptions about why a product appeared.
For background on how shopping answers and product feeds can affect product visibility, read Search Engine Land's guide to product feeds for ChatGPT Shopping, Perplexity and AI commerce. The article is useful context, but your own test results should determine what is happening for your products and target prompts.
Keep tests separate by assistant. If an interface offers different modes, such as a search-enabled mode and a standard chat mode, record them as distinct entries. If a product card is not available in a particular mode, mark the field as not applicable rather than recording a missing card as a failure.
4. Run repeated tests under consistent conditions
AI answers can vary. A prompt may receive different products on different runs, even when its wording is unchanged. One test can therefore create a misleading impression of visibility. Repeat prompts across test cycles and look for patterns across the set rather than drawing conclusions from a single response.
Use a clean and documented test environment. If you test manually, avoid adding personal preferences or past conversation context that could influence the answer. Record whether you were signed in, whether search was enabled and any other setting that could affect the result. If your team uses an API or an automated platform, log the model or mode and relevant configuration where available.
Keep the prompt wording unchanged when you are comparing periods. If you need to improve a prompt because it is unclear or irrelevant, version it and avoid comparing the revised prompt directly with the old version as if nothing changed. You can run both versions in parallel for a transition period, then use the new version as your ongoing baseline.
Choose a steady reporting cadence. Weekly or bi-weekly checks may suit a fast-changing catalogue or active campaign. A slower cadence may be enough for a stable range. Review the trend across several cycles before reacting to a short-term fluctuation. If your team needs a starting point for tracking repeatability, Siftly's AI search share of voice playbook offers further measurement guidance.
5. Record the answer in a structured log
Create one row per prompt run, not one row per reporting period. That makes it possible to inspect the underlying answer when a percentage changes. Keep the raw response or a stable record of it alongside your coded fields, subject to your organisation's privacy and data policies.
A useful log includes:
- Prompt ID, prompt wording and prompt category.
- Test date, market and language.
- Assistant, interface or mode, and account or configuration notes.
- Whether the response was valid and, if not, why.
- Brand mentioned: yes or no.
- Exact product or SKU named, where identifiable.
- Position or order in the answer.
- Whether a product card, buy link or other shoppable element appeared.
- Whether the assistant cited your domain, a competitor domain or a third-party source.
- The answer's framing, such as positive recommendation, neutral inclusion, limitation or warning.
- Notes about inaccurate details, unavailable products or price and stock claims.
Use controlled labels. For example, record positive, neutral or negative rather than allowing each reviewer to invent a new description. If a response contains a comparison, note whether the assistant names your product but recommends a competitor. A brand mention is not automatically an endorsement.
6. Apply one counting rule and calculate the metrics
Choose what counts as a mention before you calculate. For a brand-level view, count a brand no more than once per response, even if the assistant repeats its name. For a product-level view, count each distinct SKU once per response. Keep the two views separate. If the answer mentions a parent brand and a product line, state how those are handled.
Then calculate the metrics from valid runs:
- Mention rate: responses that mention your brand divided by valid responses, multiplied by 100.
- AI shopping share of voice: your counted brand or product mentions divided by all counted mentions for the defined competitor set, multiplied by 100.
- Top-pick rate: valid responses where your product meets your top-pick definition divided by valid responses, multiplied by 100.
- Card rate: valid responses showing an eligible product card divided by valid responses in interfaces where cards can appear, multiplied by 100.
For a position-weighted view, assign a higher value to an earlier recommendation and lower values to later placements. You might use a decay such as 1.0 for first position, 0.7 for second, 0.5 for third and 0.25 for later positions, as a chosen analytical model. These weights are not a universal standard. Document them and report the unweighted result beside them so readers can see the underlying counts.
Do not combine citations, mentions and product cards into one total unless you have a clear reason and a disclosed weighting method. A linked citation may support a recommendation, but it is not the same thing as a product being named. Likewise, a card can give a product useful visibility without proving that it was the assistant's preferred option.
7. Compare results by category, assistant and product
A single overall share can conceal important differences. Break the results down by prompt intent, product group, assistant, region and test period where you have enough valid observations. Look for repeated patterns: perhaps your brand appears for one use case but not another, or a particular product is named but rarely shown in a card.
Use a simple comparison table in your reporting:
| Segment | Mention rate | AI shopping share of voice | Top-pick rate | Main observation |
|---|---|---|---|---|
| Category prompts | Your result | Your result | Your result | Which products appear in broad discovery answers? |
| Comparison prompts | Your result | Your result | Your result | Are you included, and how are trade-offs framed? |
| Constraint prompts | Your result | Your result | Your result | Are key attributes recognised? |
| Use-case prompts | Your result | Your result | Your result | Does the recommendation match the shopper's context? |
Use actual values from your own log. Avoid comparing segments with very different prompt coverage as if they were equally reliable. Include the number of valid runs and any exclusions in the report, so stakeholders can understand the basis for each result.
8. Investigate the answer, not just the score
When your share rises or falls, inspect the individual responses behind the change. Check whether the assistant named a new competitor, changed the order, stopped showing a product card or cited a different source. Review the product details in the answer for errors, including price, availability, sizing, compatibility and product features.
Look for sources that recur in relevant answers. If assistants repeatedly cite an independent review, retailer listing, forum discussion or buying guide, examine what that page says about your category and products. This can reveal gaps in the information shoppers and answer systems encounter. Do not assume that one source caused a recommendation; treat recurring citations as leads for investigation.
Also check whether your own product information is complete and consistent. Product titles, descriptions, identifiers, price, availability and attributes should agree across your store and the feeds or platforms you manage. Search Engine Land's product-feed guidance is a useful reference for thinking through feed quality, while PayPal's announcement about Perplexity Instant Buy provides context on one commerce experience. Feature availability can vary, so verify current platform documentation before making implementation decisions.
9. Turn findings into specific actions
Tie each finding to an action you can test. If an assistant mentions your product but gets a key attribute wrong, review product-page copy and structured data for accuracy. If a product appears in text but not in an eligible shopping card, inspect the relevant merchant feed and platform requirements. If a competitor is repeatedly recommended for a particular need, compare the products and the information available to the assistant before deciding what to change.
Use an action log with four fields: observation, likely explanation, proposed change and next test date. Keep the explanation labelled as a hypothesis until a new test supports it. For example:
- Observation: Your running shoe is named for road running but not for wet-weather use.
- Hypothesis: The product's water-resistance attributes are missing or unclear in the information available to assistants.
- Action: Check the product page, catalogue data and reviews for accurate, consistent details.
- Next test: Re-run the same wet-weather prompts after the change and compare the recorded answers.
Do not change a page simply to repeat a phrase in a prompt. Make useful information easy to verify, keep product claims accurate and provide clear comparisons where they genuinely help shoppers. For more practical e-commerce context, see SearchFIT's article on AI recommendation pitfalls for e-commerce.
Tips for a reliable measurement programme
- Keep the prompt basket stable: Store prompt IDs and version changes. Stability is essential for comparisons over time.
- Separate discovery from brand demand: Report unbranded and branded prompts in different groups.
- Track products as well as brands: A brand may appear often while the specific product you want to sell does not.
- Preserve raw answers: A score is easier to check when reviewers can return to the response behind it.
- Label uncertainty: Record missing data, unavailable features and ambiguous product names rather than guessing.
- Use a competitor set that makes sense: Include direct alternatives, not every popular brand in the wider market.
- Pair visibility with business measures: Where possible, compare AI referral data and sales signals, but do not claim that a recommendation caused a purchase without evidence.
- Review accuracy as well as presence: A product mention with wrong details may harm trust even if it increases the score.
If you need to report across multiple clients or product lines, decide on a shared core method and allow a documented category-specific prompt set. SearchFIT's AI search visibility tracking can support ongoing visibility work; the exact workflow should match your reporting needs and the assistants you plan to monitor.
Mistakes to avoid
Treating one answer as a benchmark
A single response can be an outlier. It cannot tell you whether your products are reliably included across prompts, sessions or assistants. Repeat tests and report the scope of the sample.
Counting all appearances as equal
A first recommendation, a passing mention and a product card are different types of visibility. Report them separately or explain exactly how any combined score is weighted.
Changing the prompts between reports
A revised prompt basket can change the result even if assistant behaviour stays the same. Preserve prompt versions and note when a test is no longer comparable with its predecessor.
Treating a mention as a sale
Share of voice measures representation in tested answers. It does not directly measure shopper exposure, clicks, conversion or revenue. Use analytics and commerce data to investigate those outcomes separately.
Assuming every assistant uses the same product sources
The same prompt can produce different products and sources across systems. Do not use one platform's result as a proxy for all AI shopping visibility. Keep platform-level results in your report.
Reporting a percentage without the denominator
A percentage without its valid-run count, prompt scope and competitor set is hard to interpret. Include those details, plus exclusions and any changes to the method.
Overreacting to a short-term movement
A change may reflect variation in responses, a changed interface or an actual shift in recommendations. Check the underlying answers and look for a pattern across repeat tests before recommending a large change.
How to report AI shopping share of voice to stakeholders
Keep the main report brief and make the method easy to inspect. Start with the business question, the period covered and the assistants tested. Show the overall share alongside mention rate, top-pick rate and product-card rate, then identify the categories or products responsible for the movement.
Add a short explanation of how you counted mentions and which competitors were included. Note the number of prompts and valid runs, the reporting dates, and whether the prompts were branded or unbranded. Include examples of answers that illustrate the finding, with personal data removed if applicable.
End with a small set of actions and owners. Each action should have a reason and a way to test whether it helped. For example, a product-data fix can be followed by the same prompt tests, while a change to a comparison page can be evaluated against prompts where the assistant cited competing sources. Keep the measurement method stable while you evaluate the change.
Wrap-up: make the score useful
AI shopping share of voice is most useful as a repeatable diagnostic, not a standalone success measure. Define the category and competitor set, test realistic prompts across relevant assistants, log the answer details and use one consistent formula. Then investigate why products appear or do not appear, make a specific change and repeat the test.
A clear baseline helps your team separate a genuine visibility pattern from an interesting one-off answer. Start with a prompt basket you can maintain, report brand mentions separately from product cards and top picks, and connect changes to evidence from your own results. For a next step, explore SearchFIT pricing to see how an ongoing tracking workflow could fit your team.