How Do You Monitor Changes in What AI Says About Your Products?
How Do You Monitor Changes in What AI Says About Your Products?
Learn what brand monitoring means today, why AI platforms have changed how brands are tracked, and how to build an effective monitoring system.
Haritha Kadapa
AI-shopping monitoring should not mean repeatedly searching your brand name and recording whether it appeared.
For an ecommerce team, the useful question is more specific:
What changed in the way AI systems understand, recommend and route our products—and is that change important enough to act on?
That means establishing a repeatable baseline around the shopper questions that matter, the SKUs that should compete for them, the products and competitors that appear, the claims AI makes, the sources behind those claims and the merchant ultimately surfaced.
Then the same environment needs to be measured again.
Without that baseline, a team can see a different AI answer tomorrow and have no idea whether the business actually improved, deteriorated or simply encountered normal variation.
Start by deciding what deserves monitoring
A large ecommerce catalog can generate an almost infinite monitoring problem.
Every SKU can appear across multiple shopper intents, buying stages, markets and AI engines. Monitoring every possible question equally quickly becomes impractical.
The first step is therefore to define a stable monitoring universe.
That usually starts with commercially relevant shopper questions.
A footwear brand might monitor questions such as:
“Best waterproof trail shoes under $150.”
“Comfortable walking shoes for wide feet.”
“Best running shoes for overpronation.”
The important part is that the same questions—or a deliberately controlled version of them—are tested repeatedly.
That gives the team something comparable over time.
Gravton's current public workflow follows this logic: establish a baseline around demand and intent clusters, identify where products win or lose, make changes, and then measure what moved. Its platform explicitly separates Shopping Intelligence, Action Engine and Measurement, with measurement described as tracking what moved and feeding that result back into the system.
The exact prompt universe, cadence and sampling method still matter. A monitoring system that constantly changes the questions being tested cannot easily distinguish performance movement from measurement movement.
Monitor more than whether the brand appeared
Suppose yesterday an AI answer recommended one of your products and today it doesn't.
That is useful to know.
But it is not enough to diagnose the problem.
A meaningful ecommerce monitoring system should preserve several layers of the answer.
Layer | What changed? | Why it matters |
Product selection | Did your SKU enter or leave consideration? | Indicates recommendation movement |
Position / prominence | Did the product become more or less prominent? | Shows strength within the shortlist |
Product description | Did AI change the attributes, strengths or weaknesses it associates with the SKU? | Can expose changing product understanding |
Competitors | Did a new competitor appear or an existing one gain ground? | Shows category movement |
Sources | Did the evidence or cited domains change? | Can reveal a different information path |
Merchant | Is the product still recommended but routed to a different seller? | Separates product visibility from revenue routing |
Price / availability context | Did current commerce conditions change? | May explain recommendation or merchant movement |
These layers should not be collapsed into one generic “AI visibility” score.
A product can stay visible while the underlying commercial outcome gets worse.
For example, ChatGPT may still recommend the same product but surface a different merchant because merchant ranking can depend on factors including availability, price, quality and whether the merchant is the maker or primary seller. OpenAI also says merchant and product metadata can come from merchants or third-party providers and can change as product information changes.
Microsoft similarly notes that product details such as sizes, retailers, review summaries and price history depend on merchant information, and prices or availability can change.
So “our product is still there” is not enough.
Not every different answer is a meaningful change
This is one of the hardest parts of AI monitoring.
Generative systems can return different wording, ordering or supporting details without the underlying shopping decision materially changing.
Consider two runs.
Run A
Product X is a strong option for wet-weather trail running because of its waterproof construction and grip.
Run B
For rainy trail conditions, Product X stands out for waterproofing and traction.
The wording changed.
The commercial meaning barely did.
Now compare:
Run C
Product X is water-resistant and better suited to light rain.
That is different.
The system appears to have changed its understanding of an important product attribute.
A useful monitoring system therefore needs to distinguish surface variation from decision-relevant movement.
For ecommerce, meaningful change is more likely to involve things such as:
a SKU entering or leaving the recommendation set;
an important product attribute changing;
a competitor replacing your product;
a new objection appearing repeatedly;
a source type shifting materially;
a merchant route changing;
a major price or availability condition changing;
recommendation performance moving across a group of important shopper questions.
Minor wording changes should usually not create urgent tickets.
Establish a baseline before setting alerts
Before deciding what counts as improvement or deterioration, teams need a starting point.
For each monitored shopper question, record:
which products appeared;
which competitors appeared;
which important attributes were associated with each product;
which sources or domains supported the answer where visible;
which merchants were surfaced;
the relevant market and AI engine;
and the date of the observation.
At catalog scale, those observations can then roll up into more useful measures.
For example:
Recommendation coverage: across the shopper intents where a SKU should reasonably compete, how often is it recommended?
Competitor displacement: how often does a named competitor appear when the brand's eligible SKU does not?
Attribute consistency: are important product facts represented consistently across monitored answers?
Source movement: are the source domains or source types supporting the answer changing?
Merchant-routing movement: when the product is selected, where is the shopper being sent?
Google's current AI performance insights illustrate why time and scope matter in this type of analysis. Merchant Center allows merchants to filter AI-shopping performance by time period, product category and country, and exposes metrics such as search intent frequency and the number of products showing for those intents and attributes.
The principle is useful beyond Google's own surfaces: changes are interpretable only when the underlying scope is controlled.
Investigate what changed around the product
Imagine a brand monitors the question:
“Best everyday serum for post-acne marks under $40.”
For several weeks, its serum appears regularly alongside two competitors.
Then it disappears.
The wrong response is immediately:
“We lost AI visibility. Publish more content.”
The monitoring record should first show what else moved.
Did the product price increase above $40?
Did availability change?
Did one competitor launch a new product?
Did the AI system stop associating the serum with post-acne marks?
Did the supporting sources change?
Did the product remain recommended on ChatGPT but disappear on Gemini?
Did the SKU stay recommended but get pushed to another merchant?
Those observations determine the next investigation.
The existing Gravton source-intelligence work already makes the same distinction: a brand needs to separate a product-information problem from a source/evidence problem from a merchant-routing problem because each leads to a different action.
Monitoring is what tells the team that the state changed. Diagnosis explains why.
Track competitors as movement, not just a ranking
Competitor monitoring becomes especially useful when a new pattern appears repeatedly.
One isolated recommendation is weak evidence.
But imagine a competitor SKU that previously appeared in 10% of a monitored question set starts appearing across 45%, especially around the same attribute or use case.
That deserves investigation.
Perhaps the competitor launched a new variant. Its review volume changed. Retail distribution expanded. New editorial coverage appeared. Its product data became richer. Or AI systems simply began matching it to a broader set of shopper intents.
The monitoring layer should flag the movement without pretending it already knows the cause.
That distinction matters.
Observed: competitor recommendation presence increased.
Inferred: stronger external evidence may be contributing.
Not yet proven: a particular source or content change caused the increase.
The monitoring article should stay disciplined about those boundaries because generative-shopping systems expose outcomes more readily than they expose causal ranking logic.
Monitoring should lead to a decision
The point of monitoring is not to create more dashboards.
A useful change should eventually resolve into one of three outcomes.
No action.
The difference is minor, temporary or not commercially meaningful.
Investigate.
The movement matters, but the cause is unclear.
Act.
The movement is important and the team has enough evidence to choose an intervention.
For example:
A hero SKU disappears because its feed now reports it out of stock. Act.
A competitor suddenly appears across several high-value intents but the reason is unclear. Investigate.
An answer changes from “excellent grip” to “strong traction.” Probably no action.
This is why monitoring and prioritization should remain separate.
Monitoring asks:
What changed?
Prioritization asks:
Does this change matter enough to fix now?
Measure again after the fix
Monitoring becomes most valuable after the team makes a change.
Suppose an ecommerce team identifies inconsistent waterproof information across the PDP, feed and retailer listings and corrects it.
The same relevant shopper questions should then be rerun.
Did the product begin appearing more consistently?
Did the incorrect description disappear?
Did competitor displacement fall?
Did the underlying source mix change?
Did nothing happen?
A null result is still useful.
It may mean the diagnosis was incomplete or that the changed signal was not important enough to alter the recommendation.
Gravton's current public workflow explicitly describes this cycle as baseline → prioritize → fix → measure, with teams re-testing after changes and continuing to improve what matters.
That is the more useful interpretation of AI brand monitoring for ecommerce.
It is not listening for brand mentions.
It is maintaining a controlled view of how AI systems are representing and recommending products over time, and detecting the changes that deserve investigation.
The question is not:
“Did AI say something different today?”
It is:
“Did something change in the shopping decision that could affect which product wins, why it wins, or where the sale goes?”
Win your AI search demand universe
Understand how AI platforms are rewriting what brand monitoring means and where competitors are being tracked, cited, and recommended in ways your current tools aren't capturing.
VISIBILITY & CONTENT STRATEGY