Does product data enrichment work? How to test its effect on visibility and sales

Your team enriches a group of product pages. A month later, sales rise and an AI assistant recommends one of the products. Has the enrichment worked?

Possibly, but those two observations cannot tell you how much the change contributed. Prices, stock, campaigns, the season and the assistant's own behaviour may all have moved in the same month. A product data test is designed to separate those causes, so you know which part of the result the enrichment can claim.

The AI shopping measurement guide covers the reports that show AI traffic and visibility: GA4, Search Console and Merchant Center. This guide is about the experiment itself, so that one specific enrichment change is easier to evaluate.

Choose one specific change and the mechanism you expect it to work through. “Will publishing verified capacity and dimensions help customers evaluate this backpack range?” can be tested. “Will making products AI-ready increase sales?” cannot, because it names neither the change nor what should happen.

State which fields will change, which products are eligible and what you will observe, then date the plan. Without a dated plan it is too easy to choose, after the event, whichever metric looks best.

Publication and data quality

Did the approved values reach the live page and the destinations that support them? Were contradictions resolved? Did each variant keep its own image and offer? These checks establish whether the intended change happened at all.

Visibility and discovery

Did a defined set of shopping requests produce different brand mentions, cited pages or product recommendations? Did relevant search impressions or visits change? These observations are useful, but only with their collection conditions recorded.

Customer and commercial outcomes

Did relevant visitors engage differently, buy more often or bring in different revenue? Weigh the attribution you have, order volume and margin. A recommendation is not a sale, and a sale after publication was not necessarily caused by the enrichment.

Pick a group with reliable baseline data, steady availability and a real information gap. Where practical, compare it with similar products that will not receive the planned enrichment during the observation period.

Match on the factors that matter in your category: product type, price range, traffic, seasonality and availability. Random allocation within comparable groups strengthens the design, but small samples and effects that spill between products still limit what you can infer.

Never knowingly leave misleading or incorrect information live to protect a comparison group. Correct material errors on every affected record, and choose a different test question if that breaks the comparison.

Record why each product was included. Dropping a poor performer after seeing its result biases the conclusion. If an unexpected event makes a record unusable, document the event and show how including or excluding it changes the reading.

Suppose the test concerns verified dimensions and capacity. Keep a record of the old and new values, the evidence each was approved on, the publication time and the check that each destination received it.

If the goal is to isolate dimensions and capacity, avoid rewriting titles, changing prices, replacing photography or launching a major campaign on the same products at the same time. When another change cannot wait, log it and narrow the conclusion to match.

For Google-facing product information, Google's product structured data documentation describes the markup, and its general structured data guidelines say Google does not guarantee that correctly marked-up data will show in search results. Correct markup earns neither a rich result nor a sales lift by itself, so it belongs in the publication layer of the test, not the outcome.

Write realistic shopping requests that reflect different customer needs. For each observation, record the platform, date, exact wording, location and settings, and whether the answer contained a brand mention, a cited page or a product recommendation.

Repeat the same protocol over time, save the evidence and keep the observations where your products did not appear. Rephrasing a request until your product shows up, then counting only that answer, measures your persistence rather than the enrichment.

With a small prompt set, report counts beside percentages. “Recommended in 6 of 20 recorded observations” shows the denominator that “30% visibility” hides, and it still describes that set of observations, not a share of every shopper conversation. Sighted's pilot with The Game Collection reported its result with the set named: 200 products were enriched and 17 collections created, and AI citation coverage went from 0% to 32% of the priority products tested, and the store's organic conversion rate rose 64%.

For ChatGPT, OpenAI says the answer uses the request and its context, memory and custom instructions included, so the settings belong in the record; the ranking factors guide lists the rest of what it weighs.

Google's guidance on its AI features says that meeting its requirements does not guarantee a page will be shown. Treat a change in observed answers as evidence to investigate rather than a universal rule for being recommended.

This is an illustrative calculation, not a client result. Imagine the enriched group takes 100 orders in the period before the change and 120 in the period after, a 20% increase. A comparable group that was not enriched moves from 100 to 115 orders, a 15% increase.

The difference between those growth rates is five percentage points, so “enrichment increased orders by 20%” is not a justified reading. Even the five-point gap is not automatically caused by the enrichment: differences between the groups, stock changes and random variation still need assessing.

Check traffic and price too. More orders could come from more visits, a promotion or a different product mix, and a conversion rate built on very few sessions can swing sharply without being strong evidence.

Report the observation, the comparison and the uncertainty that remains.

Sighted's pilot with The Bean Coffee Company has the shape of that comparison: the pilot products were enriched, and their revenue rose 31% against 10% across the rest of the catalogue, with search clicks going from 121 to 366. Read with this guide's method, the rest of the catalogue is the comparison, so the reading starts from the gap between the two rates rather than the 31% on its own, with the caveats of the example above still applying.

No standard test length proves an AI visibility effect. Set the duration from purchase frequency, expected traffic, planned catalogue changes and the time each destination takes to process updates.

Fix the observation period and a review point before you start, and do not stop at the first positive week. If volume is too low for a dependable commercial conclusion, say so, and use the test to learn about data quality and the measurement process instead.

For a high-value decision that needs statistical confidence, bring in an analyst to design the test and assess the sample size before the change goes live. A more elaborate chart afterwards cannot repair a test that was underpowered or poorly controlled from the start.

Separate verified improvements, suggestive observations and open questions. Recommend whether to expand, revise or pause the programme, and name the evidence that would change that recommendation.

A test can succeed operationally while its commercial effect stays unclear. It might show that the source data improved but one publication route failed, or that the chosen prompts do not match the questions shoppers actually ask before buying. Both findings make the next decision better informed, and the next test better designed.

Can a before-and-after chart prove enrichment caused growth?

Not on its own. You also need comparison data, a record of other changes, enough volume and a reliable way of measuring.

What if you cannot keep a comparison group?

Keep a clear change log and repeat the same observations over time, then describe the findings as observational. Avoid causal claims the design cannot support.

Is a higher readiness score enough to show enrichment worked?

A higher readiness score supports a data-quality assessment. It does not establish more recommendations, more qualified visits or more sales without separate evidence of each.

1. Write down one specific change and the mechanism you expect, name the fields, the eligible products and what you will observe, and date the plan.
2. Pick a group with reliable baseline data, steady availability and a real information gap, and a comparable group that will not be enriched during the period, recording why each product is included.
3. Correct misleading or incorrect information on every affected record first, and choose a different test question if that breaks the comparison.
4. Publish the one change with a record of the old and new values, the evidence, the publication time and each destination's check, and log any other change that cannot wait.
5. Write the shopping requests in advance, record each observation's platform, date, wording, location and settings, and repeat the same protocol, keeping the answers where your products did not appear.
6. Fix the observation period and a review point before you start, then read the result as the gap between the two groups' growth rates, with traffic and price checked.
7. Report verified improvements, suggestive observations and open questions separately, and recommend whether to expand, revise or pause the programme.

Sighted scores each product on the fields shopping assistants read, drafts missing ones from your own product data and, unless you turn on automatic publishing, waits for you to approve each draft. On paid plans, on a schedule, it asks ChatGPT, Perplexity and Gemini a set of shopper questions (Sighted's suggestions until you approve your own) and records whether each answer names your brand, links your site or recommends one of your products. Choosing a fair comparison, logging other changes and deciding what a result means stay with your team.

Its Results page sets enriched products beside those not enriched, over the same two periods, on product page sessions from your Google Analytics and on sales where your store's orders reach Sighted, can leave out newly launched products, and labels any difference as observed performance, not proof of cause. The free AI visibility audit is the baseline for the first layer on up to 50 of your product pages, and the enrichment under test is the technology behind SO-ME, built to enrich catalogues at scale for conversational discovery.