Google AI Mode shopping optimization: what I fix first

What Google AI Mode actually reads before it names a product

AI Mode does not rank ten blue links and let the shopper sort it out. It answers a question, and to answer it the model has to pull named products with prices, attributes and availability out of Google's Shopping Graph. Your influence over that happens in exactly two places: the product data you send to Merchant Center, and the structured markup on your own product pages that corroborates it.

The distinction that trips up most retailers is paid versus organic. Google's own AI performance report scopes its data to organic AI traffic — free listings — and states plainly that paid Ads traffic is not included. If you run Shopping campaigns but never switched on free listings, your products are not in the organic candidate pool that AI Mode draws from, no matter how much you spend.

Everything below is work on rows one and two. There is no third lever.

The attributes that decide whether you are even a candidate

Conversational shopping queries are attribute queries wearing a sentence. "A rug for a dining room that is easy to clean" resolves to material, pile height, care instructions and room type. If those fields are empty in your feed, no amount of product copywriting puts you in the answer — the model has nothing to match against.

Google makes this unusually explicit. The AI performance report includes a popular attributes scorecard whose stated purpose is to show structured specs shoppers are asking for — size, colour, material — that may be missing from your product data sources. That is Google telling you which empty columns are costing you impressions.

The metric that matters here is coverage, not presence. A feed where material is populated on 12% of SKUs is, from the model's point of view, a feed with no material attribute at all. Audit it as a percentage per category per field, and fix the worst field across the most SKUs first — not the most interesting field on your hero product. Getting coverage from the low teens to the high nineties across a catalogue of any size is a data engineering job rather than a copywriting one, which is the part most teams underestimate; it is also most of what product feed management work consists of in practice.

Write titles and descriptions for the question, not the keyword

The classic feed title formula — brand, model, key attribute, size — is still correct. What changes is what you put in the remaining characters and in the description.

Look at the examples Google uses for its top terms scorecard: phrases like "maximum cushioning" and "arch support". Those are not product names or category keywords. They are functional benefit phrases, and they are what shoppers type when they describe a problem instead of naming a product. The optimisation loop is mechanical: pull the high-frequency terms for your category, check which ones are factually true of a given SKU, and work those into the title, description and the matching structured attribute.

Two hard rules from doing this on real catalogues:

• Never claim a benefit term the product does not have. It is a disapproval risk, and on a surface where reviews are summarised next to your product, the answer contradicts you in the same sentence it names you.
• Do not let benefit language push the identifier out. On a parts catalogue I run, the overwhelming majority of organic entries are people typing an exact part number, then narrowing by keyword. GTIN and MPN exactness is still the single highest-value field in that feed, and benefit phrases go after it, never instead of it.

One gotcha worth paying for once rather than learning yourself: bulk-generated description padding does not read as enrichment. On the same client store, the owner published a batch of AI-written content against my advice and organic rank and clicks fell by roughly 90%; recovery took months of removing and rewriting. Writing for AI surfaces and generating text with AI are different activities. Specificity is what gets you quoted. Volume is what gets you demoted.

Measure it: the AI performance report in Merchant Center

Since 2026 you can see some of this directly instead of guessing. In Merchant Center, go to Analytics → Products → AI performance. The report covers conversational queries with shopping intent on AI Mode and AI Overviews, and gives you four metrics: your share of voice, your competitors' average share, frequency (how popular a search type or attribute is), and products showing.

It classifies queries into three shopping stages — discovery, evaluation, and ready to buy — which is the part that makes it actionable. The pattern you are hunting is high frequency paired with low share of voice: a search type lots of people use, where you barely appear. That is your week's work, and it usually points straight back at a missing attribute.

Before you trust a single number in it, know these:

• Availability is narrow. English-language queries only, for Merchant Center accounts in Australia, Canada, India, New Zealand and the United States. Google's AI performance insights documentation lists the current scope.
• A share of voice of 100% is usually bad news. It most often means Merchant Center has no competitor set for your account, not that you own the category.
• 0 and "–" mean different things. Zero means insufficient impressions; a dash means no impressions data at all.
• You cannot pick your competitors. The benchmark set is the one Merchant Center defines for you.
• Your share and your competitors' average can both rise at once when that set changes. Read the trend, not the absolute.
• There is no all-categories view and data lags by a few days. Compare like for like, and do not react to a single day.

Also note the traffic filter is locked to organic. This report will never explain the performance of your Shopping campaigns, and that is deliberate.

What to do if the report is not available in your market

Five countries is not most of the internet. If you sell from outside that list, you still need a feedback loop, and the honest answer is that you have to assemble proxies and label them as proxies.

1. A fixed prompt panel. Write 20–30 conversational prompts that represent how buyers actually describe your products, covering all three stages from discovery to ready-to-buy. Run them monthly, from a clean session, in the markets you sell to, and record one thing: were you named, and in what terms. It is manual and it is the only method that shows you the actual answer text rather than a metric about it.
2. Free listings performance by product in Merchant Center, read as a directional organic signal rather than an AI-specific one.
3. Search Console impressions against clicks on product URLs. Impressions holding or rising while clicks fall is the classic signature of being summarised rather than visited.
4. A manual attribute gap audit against three competitors you would want to be benchmarked against. You are reproducing the popular-attributes scorecard by hand, and it works.

Do not present any of these to a client as share of voice. They tell you direction, not position.

Keep the product page and the feed telling the same story

Schema.org Product and Offer markup is corroboration, not a substitute for the feed. Price, availability, condition, GTIN, MPN and aggregate rating should match the feed exactly, because a disagreement is read as unreliability on both sides, and in the worst case it is a disapproval.

Three practical checks that catch almost everything:

• Generate JSON-LD from the same source of truth that builds the feed. Hand-maintained markup drifts within a quarter — always.
• Refresh the feed at least as often as prices change. If you reprice intraday, push through the Content API rather than waiting on a scheduled fetch.
• Spot check availability on a random sample of twenty URLs weekly, comparing page, markup and feed. The mismatch is almost never on the product you would have guessed.

Agentic checkout, Business Agent and Direct Offers: what to act on now

On 11 January 2026 Google launched the Universal Commerce Protocol (UCP), an open standard for agentic commerce spanning discovery through post-purchase support. It was co-developed with Shopify, Etsy, Wayfair, Target and Walmart, endorsed by more than twenty others including Stripe, Visa, Mastercard, Adyen, Best Buy and The Home Depot, and it is compatible with A2A, AP2 and MCP. Its first application is a checkout feature on eligible product listings in AI Mode and the Gemini app, for eligible US retailers, paid with Google Pay using details already in Google Wallet — with the retailer remaining seller of record. The same announcement introduced Business Agent, a branded chat agent on Search that eligible US retailers activate in Merchant Center, and Direct Offers, an Ads pilot for surfacing exclusive discounts inside AI Mode.

That is a lot of surface area, and almost none of it is where your next hour should go.

My read: the protocols will keep moving, and a protocol you integrated in 2026 may be a migration in 2027. The structured product data underneath does not move. A catalogue with complete, true, machine-readable attributes is the asset that survives every rename, and it is the only one that pays off on Google, on Gemini, and on whatever surface reads a feed next year.

The order I would fix things in

1. Enable free Shopping listings and clear Diagnostics disapprovals. Nothing downstream matters while products are ineligible.
2. Measure attribute coverage as a percentage per field, per category. Fix the weakest field across the most SKUs first.
3. Rewrite titles for your top fifty revenue SKUs with true benefit language, identifiers intact.
4. Align schema with the feed, generated from one source.
5. Raise refresh frequency, or move price and stock to the Content API.
6. Stand up the measurement loop — the AI performance report if you can see it, the prompt panel if you cannot.
7. Start collecting Q&A, compatibility and substitute data as fields rather than paragraphs, ahead of the attribute rollout.

Steps one, two and four are where the gains are, and they are also the three nobody can see from the outside — which is why most stores do not know they are failing them. If you want the gap list for your own catalogue without doing the audit yourself, the $20 Audit + Roadmap returns a prioritised fix list in 24–48 hours.