How AI Shopping Agents Evaluate Your Product Listings (Step by Step)
Most ecommerce managers understand, in the abstract, that AI shopping agents are becoming important. What most do not have is a precise model of what happens inside the evaluation, the specific steps an agent takes between receiving a shopper’s intent and deciding which products to recommend. That precision is what this article provides. Because once you understand the mechanism, the data gaps that are costing you agent visibility become obvious, and the fixes become targeted rather than generic.
The 5-Stage Evaluation Flow
All AI shopping agents, regardless of whether they are embedded in Google, Amazon, a browser, or a third-party shopping tool, execute the same fundamental evaluation sequence. The inputs vary by system. The logic is structurally identical.
Intent parsing
Natural language becomes structured criteria.
Data retrieval
The agent hits product data sources and builds a candidate set.
Structured matching
Each candidate is evaluated against binary criteria.
Ranking
Only products that pass every criterion are sorted.
Action
The agent recommends, buys, or re-queries.
Intent Parsing — Natural language becomes structured criteria
The shopper’s request (“find me a packable waterproof hiking jacket under 500g, under £150, good reviews, ships fast”) is parsed by the agent’s natural language understanding layer. The output is not a search query. It is a structured set of evaluable criteria: category = hiking jacket; packable = true; waterproof = true; weight < 500g; price < £150; rating >= 4.0; shipping_days <= 3. This parsing step is where the shopper’s qualitative language is translated into the binary filter logic the agent will apply to product data.
Data Retrieval — The agent hits product data sources
The agent executes retrieval against one or more data sources: Google’s Shopping Graph API, Amazon’s Product Advertising API, schema.org markup on product pages, or a retailer’s own product API if accessible. This retrieval step returns a set of product records that are candidates for evaluation. The scope of what is retrieved depends on how well your product data is indexed on the relevant platform. GTIN entity matching, taxonomy classification depth, and feed completeness all affect whether your product is in the candidate set at all.
Structured Matching — Binary criterion evaluation
Each candidate product record is evaluated against the structured criteria from Step 1. This is not a relevance score. It is a pass/fail test for each criterion. A product either has waterproof = TRUE in a structured field (pass) or it does not (fail). A product either has weight < 500g in a numeric attribute field (pass) or it does not have a weight attribute at all (automatic fail, not a partial pass). Every criterion that fails is an exclusion. Products that fail any single criterion are removed from the consideration set entirely.
Ranking — Sorting the products that pass all criteria
Products that survive the binary matching stage are ranked by the agent using a weighted combination of factors: review score and volume, price competitiveness within the category, shipping speed reliability, merchant trust signals accumulated over time, and attribute specificity (products with more precise, detailed attribute data are ranked higher as higher-confidence matches). This ranking stage is where enrichment quality creates a secondary advantage. Even among products that all pass the binary criteria, richer data ranks higher.
Action — Purchase, recommendation, or re-query
The agent takes one of three actions: it purchases the top-ranked product (full autonomy), it surfaces a ranked shortlist for the shopper to choose from (semi-autonomous), or it re-queries with relaxed criteria if no product passed all criteria (adaptive search). Your product must pass the binary matching stage to have any chance of appearing at the action stage. No amount of ranking optimization matters if binary exclusion removes you from the consideration set.
The No-Partial-Credit Rule — Explained at the Mechanism Level
This is the most commercially significant behavioral difference between agent evaluation and keyword search, and it deserves a precise explanation.
In a keyword search ranking algorithm, relevance scores are continuous variables. A product with “waterproof” in its description scores lower on waterproof relevance than a product with waterproof = TRUE in a structured field, but it still receives a non-zero score and appears in results, perhaps at position 12 instead of position 3. The algorithm treats missing or weak data as a ranking signal, not a disqualification trigger.
In an agent’s structured matching layer, criteria are binary predicates. The query is:
A product without a weight attribute does not receive a weight score of “probably under 500g based on the description.” It receives a NULL value, which evaluates to false for the predicate weight < 500. The product is excluded. Not downranked. Excluded.
Why NULL Is Worse Than Wrong
A product with weight = 620g fails the weight < 500g criterion and is excluded, but at least the data is present and accurate. A product with no weight attribute also fails, but additionally has no data for the agent to work with on any future query involving weight. NULL values are a double loss: they cause current exclusion and they provide no signal for the agent to learn from. Absent attributes are structurally worse than incorrect attributes, because incorrect attributes can at least be corrected and will match some queries.
How Agents Rank Products That Pass All Criteria
Passing the binary filter stage is the prerequisite. Ranking determines your position among products that have passed. Understanding the ranking factors explains the second layer of competitive advantage that enrichment quality creates.
| Ranking Factor | Mechanism | Data Enrichment Action |
|---|---|---|
| Review score × volume | The agent weights review score (4.3 is better than 4.0) multiplied by review volume (4.3 from 2,400 reviews outranks 4.5 from 12 reviews). This is a trust signal the agent can evaluate quantitatively. | Implement aggregateRating schema to expose review data to crawling agents; GTIN entity matching consolidates cross-seller reviews in Google Shopping Graph. |
| Price competitiveness | The agent compares your price against the median and range of other products in the retrieved candidate set. Being priced at the category median for equivalent attributes ranks higher than being a price outlier in either direction. | Ensure real-time price accuracy in feeds and schema; accurate sale_price and promotional pricing data; price consistency across channels agents might compare. |
| Attribute specificity | Products with more precise, unit-based attribute values are evaluated as higher-confidence matches and ranked higher. A product with weight: 490g ranks above one with weight: approximately 500g, which ranks above one with no weight attribute. | Replace all descriptive attribute values with precise, unit-based equivalents; every numeric attribute should have a specific value with unit. |
| Shipping confidence | Agents optimizing for shopper satisfaction weight shipping reliability highly. Products with specific delivery time commitments (shipping_days: 2) rank above those with vague commitments (“fast shipping”). | Populate shippingDetails schema with specific transitTime min/max values; keep shipping data current and accurate; Prime or guaranteed-delivery signals improve rank. |
| Merchant trust accumulation | As agents interact with merchants over time, they develop trust signals based on data accuracy, fulfillment reliability, and return rates. High-trust merchants rank higher even when attribute data is otherwise equivalent. | Maintain price and availability accuracy; reduce “not as described” return rates through attribute accuracy; consistent data quality is a long-term competitive moat. |
What Agents Cannot Be Persuaded By
This is as important as understanding what agents respond to. The following things that work for human shoppers and keyword algorithms have zero or negative effect on agent evaluation:
- Marketing language, such as “industry-leading,” “premium quality,” and “best in class.” Agents do not parse qualitative claims. They match attribute values. Marketing copy is invisible to the evaluation logic.
- Brand storytelling, including the origin story, the brand mission, and the founder’s journey. Agents are not evaluating brand resonance. They are matching criteria. Well-written brand narrative has no effect on agent ranking.
- High-quality lifestyle imagery. Agents do not currently evaluate image quality in Shopping Graph queries. Images matter for conversion after the agent recommendation, but they do not affect the agent’s product selection.
- Persuasive description copy. A beautifully written 300-word description that does not contain structured attribute values contributes minimal signal to an agent evaluation. The agent reads structured fields, not prose.
- Social media presence and external brand signals. Agents querying product APIs do not have access to your brand’s social media engagement, follower counts, or influencer associations. These signals do not enter the evaluation pipeline.
The Copywriting Investment That Will Not Pay Off Agentically
This is not an argument that copywriting does not matter. It absolutely matters for human shoppers who read your PDP, for conversion rates after an agent recommendation, and for semantic matching in hybrid systems like Amazon Rufus that combine structured and semantic evaluation. The point is narrower: investing in better marketing copy at the expense of structured attribute completeness is the wrong tradeoff for the agentic era. Attributes first. Copy second.
A Worked Example: The Same Product, Two Data States
Here is a concrete walkthrough of how an agent processes two versions of the same product, one with sparse data, one fully enriched, for the query “packable waterproof hiking jacket under 500g, under £150, 4+ stars”:
| Evaluation Criterion | Sparse Product Data | Enriched Product Data |
|---|---|---|
| packable = TRUE | FAIL — “packable” mentioned in description prose but no structured boolean attribute. Agent filter returns NULL. | PASS — packable: true in dedicated attribute field. Immediate match. |
| waterproof = TRUE | FAIL — “waterproof design” in description. No waterproof attribute field. NULL result. | PASS — waterproof: true; waterproof_rating: 20000mm HH. Both structured. |
| weight < 500g | FAIL — “extremely lightweight” in description. No weight attribute. NULL result. | PASS — weight: 490g. Numeric, unit-based, directly comparable. |
| price < £150 | PASS — price field is populated correctly. | PASS — price field populated; sale_price populated for current promotion. |
| rating >= 4.0 | PASS — review data present. | PASS — rating: 4.3 from 847 reviews; aggregateRating schema confirmed. |
| Overall result | EXCLUDED — fails 3 of 5 criteria due to missing structured attributes. Product is invisible to this agent query despite being a perfect physical match. | INCLUDED and ranked highly — passes all criteria; attribute specificity, review volume, and shipping data contribute to high ranking position. |
The Action Step: What This Means for Your Catalog Right Now
The worked example above illustrates a product that is a perfect physical match for the shopper’s requirements, is competitively priced, and has strong reviews, and is still completely invisible to the agent query because of three missing structured attributes. That product is losing sales today, in live agentic systems, because of data gaps that could each be fixed in a single data update.
The immediate priority list for any ecommerce manager reading this:
- Identify your category’s top purchase criteria — what are the 8–10 attributes shoppers most commonly specify when buying in your category? These are your agent filter requirements.
- Audit structured field coverage for those criteria — not description coverage. Structured attribute field coverage. For each criterion, what percentage of your SKUs have it populated as a typed field?
- Fill the gaps for your highest-revenue products first — start with the products where agent exclusion costs the most. Each missing attribute on a high-revenue SKU is a daily revenue gap you can quantify.
- Replace descriptive values with precise ones — “lightweight” → “490g”; “water-resistant” → “waterproof: true, rating: 15000mm HH”; “fast shipping” → “shipping_days: 2”.
Start with intent scenarios
The attributes shoppers specify are the filters agents are most likely to apply.
Measure field coverage
Description mentions do not count. Typed structured fields do.
Fix the highest-value gaps first
A missing field on a high-revenue SKU is an active visibility loss every day.
Velou on Simulating Agent Queries Against Your Catalog
One of the most valuable things you can do before investing in enrichment is to simulate agent queries against your own catalog. Pick your top 10 purchase intent scenarios for your primary category and run each as a structured attribute filter against your product database. Count how many products survive all filters.
For most retailers, the first time they do this exercise, the number is sobering, and the direct line from data gap to commercial consequence becomes impossible to ignore. Commerce-1 includes a catalog simulation tool that automates this exercise across your full product range.
Simulate how AI agents evaluate your catalog today
Commerce-1 runs agentic query simulations against your product data, showing exactly what agents see and what they miss.
Request a demo
.png)
.png)