Google is testing AI-written Shopping ad copy. Bing and ChatGPT Ads are testing whether they need your copy at all.

Within six days of each other in late July 2026, three shopping platforms tested a version of the same idea: render the ad or listing straight from the product feed instead of from whatever the merchant’s own copy says. Google is testing AI-generated descriptions on Shopping and Product ads.¹ On Bing, the test prints prices directly onto the product image in organic search results.² ChatGPT Ads builds its own product-feed ad units by pulling title and description straight from the feed, with a separate July update adding refreshed cards for pricing and star ratings.³⁴ Practitioners spotted two of the three independently. Currently, none is a confirmed launch. In every case, what holds is the direction: control over the final ad is shifting from the merchant’s copy toward the feed.

Three shopping surfaces tested feed-rendered content within six days of each other in late July 2026. None has confirmed shipping it.

What is Google testing in Shopping and Product ads?

As Search Engine Land first reported, Google is testing AI-generated descriptions that appear alongside Shopping and Product ads, extending a similar experiment it already confirmed on standard Search ads earlier in July 2026.¹ A paid-search practitioner spotted the extension to Shopping surfaces on July 30.

Google has described only the earlier Search-ads version, calling it a small experiment to see whether adding AI-generated context helps people make more informed decisions, and has not said whether either test will expand beyond its current scope.¹ Because Google has not said what feeds the generated text, a merchant cannot tell whether the copy draws on feed attributes, on-page prose, some crawled combination, or something the company has not described at all.¹ That silence matters more than the test itself. A merchant currently has no visibility into copy that could appear on their own ad, and no way to opt out once it runs.

What is Bing testing on organic product images?

Where review stars normally sit on a Bing product image, the company is testing a price instead, printed directly onto the image inside organic search results. A paid-search practitioner posted a screenshot of the change on July 29, 2026, and Search Engine Roundtable confirmed seeing the same pattern, plus a second and apparently unrelated test showing two rows of product images stacked under a single search snippet.²

Microsoft has not commented on either change, and this one sits in Bing’s free organic listings, not in a paid ad. Because nothing here involves a generative model, Bing is simply reading a price field that already exists in its product data and rendering it onto an image asset that used to carry only stars. Who wrote it matters less now than where the number lives.

What does ChatGPT Ads actually pull from a product feed?

According to OpenAI, ChatGPT Ads builds its product-feed ad units directly from the advertiser’s own feed. Its help documentation says plainly that in a feed-based campaign “the ad title and description will be pulled directly from your product feed,” and that ads built this way have been among the platform’s strongest performers to date.³

In a separate July 24, 2026 platform update, reported by Search Engine Land, the company added refreshed product-feed ad cards showing pricing and star ratings alongside conversion-based bidding, geographic exclusions, and bulk campaign tools.⁴ This signal alone needs no inference. No rewrite stands between the feed and the ad; the feed’s title and description fields are the ad copy, verbatim, by OpenAI’s own account. If those fields are thin, the ad is thin. No rewriter sits in between to compensate for what the feed itself never said.

Are these three tests one shift, or three different mechanisms?

Not quite one mechanism. Google’s AI-description test, Bing’s price-on-image test, and ChatGPT Ads’ feed pull-through all point the same direction without sharing one method. Google’s test inserts a generative model between the feed and the shopper, producing text nobody explicitly wrote.¹ In Bing’s price-on-image test and ChatGPT Ads’ feed pull-through, no model appears at all.

Both simply surface a feed attribute, a price or a title, into a spot a merchant used to fill with a separate asset or a hand-written line.²³ The label “AI rewriting your copy” is accurate for exactly one of the three: Google’s. Across all three, the shared thread is the feed’s fields, rather than the merchant’s separately authored creative, deciding what a shopper sees, model or no model.

Why would ad performance shift from phrasing to structured feed attributes?

The reasonable inference, not a confirmed mechanic, is that a rewriter or renderer sitting on top of a feed reads its structured fields more reliably than a merchant’s free-form prose, because those fields exist for exactly that purpose. Google’s own product data specification says the attributes it collects exist “to match your products to the right queries.”⁵

Because Google has not said what grounds its description generator, treat this as my read of how these systems behave elsewhere, not a confirmed fact about the test itself.

In a related system, Google’s Merchant Center already treats structured feed data as more authoritative than a merchant’s page copy: its Automations feature updates a live listing’s price, availability, condition, and even its images to match the feed and the merchant’s own website, without a merchant approving each individual change.⁶ Treat that as falling short of proving the description test works identically. It confirms something narrower: Google is already comfortable letting machine-read data override or supplement merchant-authored fields elsewhere in the same product, and that narrower fact is what makes the inference above reasonable rather than speculative.

What should a merchant do differently starting now?

Nothing changes today, since none of the three tests has shipped broadly, but the response should start before any of them does. When a platform layer renders or rewrites an ad, what earns performance shifts from the phrasing itself to the structured attributes a rewriter or renderer would need to read: brand, GTIN or MPN, size, color, material, and category.⁵

Once a platform can write, or simply copy, the title itself, writing a better one stops being the whole job.

I have run a title and description rewrite pipeline across a catalog of several million SKUs, and that pattern is the transferable finding from the work, not a claim about any single platform’s internals. The fields that mattered most were rarely the adjectives; more often they were the identifiers and variant attributes that told the downstream system, Google’s or anyone else’s, which product was actually being described.

How do you count your own feed’s attribute completeness?

Pull your live product feed export and count non-blank values in seven columns: brand, gtin, mpn, google_product_category or product_type, color, size, and material.⁵ Google’s specification marks brand as required for nearly all new products, GTIN as strongly recommended even where it is not strictly required, and color, size, and material as required specifically where they distinguish one variant from another.

A low fill rate on any of the seven is a gap a rewriter, human or model, has nothing to read.⁵

Run the same count against a ChatGPT Ads feed if you have one live. OpenAI’s own feed guidance for commerce integrations flags missing required fields, outdated or non-spec field names, and malformed field values as the ingestion failures to watch for, the identical complaint in different vocabulary.⁷ Rendering ads directly from a feed does not fix a feed that fails that count today. Only the gap’s visibility changes. It surfaces faster, to an audience the merchant no longer fully controls.