AI Mode Shows 95% Fewer Products Than Standard Search
Productrise published a study in July 2026 covering 21 days of US and UK search data: over 100,000 search results and more than 2 million product listings. Running the same queries through standard search and AI Mode, the study found roughly 95% fewer product listings in AI Mode.
These are third-party observations, not figures Google published. The sample size is large enough that the direction holds even if the exact percentages shift.
Two separate contractions, multiplied together
The 95% is not one thing. It is two independent drops stacking.
The first is whether products appear at all. Standard search surfaced product results on 88% of queries, counting shopping carousels, PLAs and product packs. AI Mode did so on 23%. Three out of four queries in AI Mode get a text answer with no commercial inventory attached.
The second is how many products appear when they do. Standard search averaged 22.5 listings per result page. AI Mode averaged 4.3. That number is the one worth internalizing, because it changes what competition means. Fighting for one of 22 slots and fighting for one of four are different sports.
| Metric | Standard search | AI Mode | Delta |
|---|---|---|---|
| Queries surfacing products | 88% | 23% | -74% |
| Average listings per page | 22.5 | 4.3 | -81% |
| Combined product exposure | Baseline | About 5% | About -95% |
| Product overlap with standard search | n/a | 0.8% | Effectively none |
Multiply the two: 0.23 times 4.3 against 0.88 times 22.5 leaves you with roughly 5% of the exposure surface. That is where the 95% comes from.
If you sell in dense categories such as home goods, phone accessories, or apparel basics, this hits harder. Those categories relied on long-tail queries and sheer listing volume to catch spillover traffic. There is no spillover in a four-slot result.
There is a second-order effect buried in the 23% figure: which queries carry commercial inventory has changed, not just how much. Explicit transactional phrasing still reliably pulls products. A meaningful share of comparison and use-case queries in the study got answered in prose alone. If your keyword strategy leans on “best X for Y” phrasing, audit which of those terms still return products at all before you optimize anything else.
The 0.8% overlap is the part that should worry you
Fewer slots would be a familiar problem. Raise bids, tighten titles, sharpen pricing. The 0.8% overlap says it is not that kind of problem.
Products that rank well in standard search appear in AI Mode at close to random rates. The two interfaces surface almost entirely disjoint sets of products. A well-optimized Shopping campaign and a strong organic position tell you nothing about whether AI Mode will recommend you.
The mechanics explain it. Standard search ranking weighs keyword match, click-through history, bid, and landing page signals. AI Mode parses intent first, then looks for products whose structured data can answer that intent. When someone asks for a folding drying rack that fits a narrow balcony, the model needs folded dimensions, load capacity, and rust resistance. None of that lives in your product title. It lives in feed attributes you probably left blank.
So stop reading Search Console rank reports as a proxy for AI Mode performance. They are not correlated. To know where you stand, run your core queries manually or use an AI visibility tracker like Profound or Peec AI to batch them.
The overlap figure cuts the other way too. If your products sit on page three of standard search, that says nothing about your AI Mode odds. A smaller brand with a genuinely complete feed can outrank an established competitor whose catalog carries only required attributes. This is one of the few situations where data discipline beats domain authority.
Feed attribute completeness is the entry gate
This is the cheapest work on the list and most catalogs have not done it. No developer, no budget line, just someone in Merchant Center for two days.
A typical feed carries only the attributes Google requires: title, description, price, image_link, availability, GTIN. That is enough to run Shopping ads. It is not enough for a language model to reason about fit. From that record, the model knows you sell a drying rack for 29 dollars. It cannot tell whether the rack suits a narrow balcony.
The optional attributes that matter: material, size, size_type, color, pattern, age_group, gender, and the dimension set product_length, product_width, product_height, product_weight. Beyond those, product_detail accepts custom section_name / attribute_name / attribute_value triples, which is effectively a spec sheet you hand to the model. Folded dimensions go here. So do maximum load, assembly time, and whatever the intended use case is.
Do not leave product_highlight empty either. It takes up to ten short benefit statements, and when AI Mode writes a recommendation rationale, that field is a likely source. Write them the way a buyer would ask. “Folds to 8cm and slides behind a sofa” is usable. “Thoughtfully designed to save space” is an adjective the model cannot match against any specific intent.
While you are in there, verify google_product_category and product_type. The first should use the deepest applicable node in Google’s taxonomy, not a second-level shortcut. The second should mirror your own store’s category path. These two fields are the most direct input into what the model thinks your product is, and a shallow value lands you in the wrong candidate pool entirely.
One detail that trips people up: several attributes are enum-constrained. size_type only accepts regular, petite, plus, big and tall, or maternity. Writing “standard” gets the value silently dropped. Same story with age_group and condition. Validate against the spec rather than assuming free text works.
Once you have pushed the changes, open Merchant Center Diagnostics and clear the item-level warnings, not just the errors. Warnings never blocked serving, so most teams ignored them. Each one now maps to information the model cannot read.
Structured data and trust signals decide the shortlist
The feed gets you into the candidate pool. On-site signals decide whether you come out of it.
Product schema needs to be complete, not present: offers with priceValidUntil, shippingDetails, and hasMerchantReturnPolicy, plus aggregateRating, review, brand, sku, and gtin13. Shipping and returns matter more than they used to, because AI Mode routinely writes “free shipping” or “30-day returns” into its answer. Without markup, it will not say it on your behalf.
Review volume also carries weight, measured per SKU rather than per store. A store with 4,000 reviews spread across 900 products looks thin at the item level. Shopify merchants using Judge.me or Loox should confirm the Google Shopping review sync is switched on so ratings flow into Merchant Center, not just the storefront widget.
Brand credibility signals are doing real work here too: a genuine About and Contact page, a published returns policy, Organization schema with sameAs pointing at your social profiles, third-party mentions of your brand name. These barely moved organic rankings. They do factor into whether a model treats you as safe to recommend.
The last one is easy to overlook. Many product pages carry two sentences of copy and rely on images for everything else. Models do not read text baked into images. Write 300 to 500 words covering material, dimensions, use case, and how the item differs from close alternatives, broken up with subheadings.
Sequencing the work
Ordered by return on effort. Week one, fix the feed. Take your top 50 revenue SKUs, fill every optional attribute listed above, build out product_detail, and max out product_highlight at ten entries each. This is Merchant Center work only, no code, and a couple of days of focused effort.
Week two, close the schema gaps, prioritizing shippingDetails and hasMerchantReturnPolicy. If you are on a Shopify theme with built-in JSON-LD, those two fields are almost certainly missing and you will need a schema app or a theme edit.
Week three, start measuring. Pick 20 to 30 commercial queries and check AI Mode weekly. Log whether you appear, in which position, and which competitor took the slot when you did not. The baseline is worthless for two months and irreplaceable after that.
On budget: Performance Max and AI Max can reach some paid inventory inside AI Mode, but organic slots are earned entirely through data quality. Of the 4.3 available positions, paid placement typically accounts for one or two. Spending more does not create positions that are not there.
阅读本文中文版: AI Mode 里的商品比普通搜索少 95%：跨境卖家怎么抢剩下的位置
Related Articles
Shopify Turned UCP On For You: Now Manage Exclusions and Field Quality
Shopify Editions Summer '26 shipped on June 17, 2026 with 150+ updates, and one of them flipped Universal Commerce Protocol on by default for every store. Eligible products now syndicate to ChatGPT, Microsoft Copilot, Google AI Mode, Gemini and the Shop app with zero configuration. The merchant question is no longer whether to join agentic commerce, but which SKUs to pull out and whether your fields are machine-readable.
Amazon's 75-Character Title Limit: Restructure Your Listings Before AI Does It for You
Amazon begins enforcing a 75-character title limit on July 27, 2026, and over-limit titles reportedly get rewritten by AI. Here is the title formula that works, what to move into the new Item Highlights field, and a five-step compliance audit.
Conversational Discovery Ads in AI Mode: How to Feed Gemini So It Builds Accurate Creative
Google Marketing Live 2026 launched two AI Mode ad formats — Conversational Discovery ads, where Gemini builds creative for each shopper's question, and Highlighted Answers, where strong ads enter recommendation lists. Both carry a Sponsored label. Accuracy depends on the structured product data in your feed.