Table of Contents
TL;DR
- After ChatGPT’s May 7 2026 Branded Link Update, Profound measured B2B software referrals lifting more than 200% above baseline while e-commerce and retail were essentially flat. The simplest reading is that ChatGPT recommends companies, not products.
- That does not mean product copy stops mattering. It means the AI shopping agent (a separate reader from the AI chat) becomes the second audience on every product page, and the page must read clean to both.
- The human reader still wants the same things they always wanted. A clear name, a real price, a credible promise, a calm CTA. None of that changes.
- The agent reader wants different things. Visible HTML pricing it can parse without running a script. Clean structured fields it can match against the buyer’s question. No "call for quote" walls. No hidden inventory state.
- The rewrite is small, and it is mostly about removing things. A page that fights the agent fights the human too, because both readers respond to the same kind of clarity.
You open one of your product pages and read it the way a buyer would. Hero image, name, price, three bullet points, an add-to-cart button. The page looks fine.
Now read the same page the way an AI shopping agent reads it. The agent does not see the hero image, it sees the HTML.
If the price arrived through a script that ran in the browser, the agent did not get the price. If the bullets sit inside a custom component the agent does not recognise, the agent did not get the bullets.
After ChatGPT’s May 7 2026 Branded Link Update, Profound measured B2B software brands lifting their daily ChatGPT referrals more than 200% above baseline, while e-commerce and retail saw essentially no lift at all. The simplest reading is that ChatGPT recommends companies, not products. The companies that gained were the ones whose homepages confirmed the recommendation cleanly. The products that did not gain are still discovered by a different surface, and that surface is the AI shopping agent that browses on the buyer’s behalf.
So the product-page rewrite is not optional. It is the move that decides whether the buyer’s agent puts your product on the short list or skips it for a competitor whose page the agent could read.
Who is the second reader on your product page?
The second reader is a software agent acting for a person. They run inside ChatGPT, inside Perplexity Comet, inside Gemini in Chrome, or inside a buyer’s own scripted workflow. They do not browse the way a person browses.
They land on a page, fetch the HTML, parse what they can, and compare what they found against the question their owner asked. The owner said "find me three lightweight backpacks under sixty dollars that ship to my address by Friday". The agent does not read your brand story. It reads the price, the weight, the shipping fields, and the stock state.
The agent’s reading is shallow and structural. It does not infer meaning the way a person does. If your price lives in a JavaScript-rendered widget, the agent does not learn the price. If your shipping ETA sits inside a modal the agent has to click to open, the agent does not learn the ETA.
That shallowness is not a flaw to design around. It is the constraint to write to. The same constraint that helps the agent (clear fields, plain values, predictable structure) helps the human buyer too, because human buyers scan first and read second.
For more on how the agent reads the brand-level story above the product page, see how to get AI to mention your brand by name.
Why did e-commerce sit out the May 2026 referral lift?
The data tells a clean story. After May 7 2026, B2B Software and SaaS daily referrals lifted more than 200% above their pre-update baseline. Financial Services lifted about 60%. E-commerce and retail lifted essentially zero.
The pattern is consistent with ChatGPT recommending companies for the kinds of questions that have a small number of right answers. "Which CRM should a five-person agency use?" returns three names. "Which lightweight backpack should I buy?" returns a category, not three names, because product-level recommendations are too dense for a chat answer and too brittle without inventory and pricing the engine cannot easily verify.
So the chat surface skipped products. The traffic went to homepages, and Similarweb measured the homepage share of OpenAI referrals rising from 26-32% before May 7 to about 60% after. The product-detail page sat outside that lift.
But the same buyer who asked ChatGPT for a CRM also asks an agent (or an agentic browser) to do the product comparison once the company short list is set. The agent fetches your product page, parses the fields, and reports back to the buyer. That second step is the product-copy job in 2026.
The chat surface decides which companies make the short list. The agent surface decides which products inside those companies make the buy. Two surfaces, two reading styles, two copy disciplines.
What does an AI shopping agent need that a human buyer does not?
The agent needs five things the human takes for granted. A price the page renders as text, not as a value injected by a script after page load. A clear product name that matches the way the agent expects the product to be referred to. A visible variant grid (sizes, colours, SKUs) without a modal trap.
A shipping field the agent can read without filling a form. An inventory state that says "in stock" or "out of stock" in plain words rather than greyed-out controls. None of those are persuasion moves. They are accessibility moves for the second reader.
The human buyer reads through the same fields without noticing them. If the price is visible, they read it. If the size selector is one tap away, they tap it. The agent reads more carefully and gives up faster.
The cost of failing the agent is not zero. The buyer who asked the agent to shortlist three backpacks under sixty dollars does not see the page where the agent could not parse the price. The agent skipped it.
That skip is invisible to the analytics. The page that fails the agent never registers a visit because the agent never opened the browser tab.
Which product-copy moves keep the human and serve the agent?
Three moves do most of the work, in order.
Write the hero block as a definition, not as a slogan. One sentence on what the product is, one sentence on who it is for, one sentence on what the buyer otherwise would have used. The agent reads the definition and matches it against the buyer’s question. The human reads it and decides whether to scroll.
Place the price, the shipping ETA, and the stock state as visible HTML in the first viewport. Not as a script-rendered widget, and not behind a modal.
The agent finds the three values the buyer asked about and reports them back. The human confirms them in one glance.
Move the proof to the right place. A short proof block (one quantified outcome, one named customer kind, one specific use case) sits in the supporting paragraph under the hero. The agent extracts the outcome as a cite-worthy claim. The human reads it as ground truth.
The proof-placement discipline that decides exactly where each block lands sits in the proof placement landing-page discipline, and the upstream homepage question that the agent has already asked sits in how to write a homepage for visitors AI sends you.
What does structured product data look like in plain words?
Structured data sounds like a developer concern. It is also a copy concern. The product page that the agent reads cleanly is the page whose fields the developer marked up as a product (name, brand, price, availability, sku, image, description) and whose copy filled those fields with the right specific values.
Name is the canonical product name, written the way the buyer would type it. Not the marketing name only. Not the SKU only.
Brand is the company name as the buyer says it out loud. Consistent across the catalogue. The agent reads the brand field and links it back to the company name in the chat answer.
Price is a number in the local currency, present as text, not as an image or a script-rendered widget. The agent reads the price and reports it to the buyer without ambiguity.
Availability is a small enum (in stock, out of stock, preorder, discontinued). The agent reads availability and decides whether to put the product on the short list at all. The human reads it and decides whether to click add-to-cart.
Description is the prose the page already had, with two small changes. The first sentence is the definition, not the brand vignette. The buyer-vocabulary words sit in the body rather than in a hidden meta tag.
Other questions worth answering
Does machine-readable HTML beat a script-rendered widget when your store runs on WooCommerce?
Yes. The shopping bot fetches raw markup before any JavaScript runs. If WooCommerce injects the dollar value through a checkout script, the bot misses it.
Render the value, the variant grid, and the stock state as static markup in the initial response. The buyer never notices the difference, and the bot finds every field on the same fetch.
Where should brand storytelling go once a bot becomes the audience?
Push the vignette below the supporting paragraph, not above the definition. The bot extracts whatever sits closest to the title field, so a poetic intro hides the canonical noun. Per Profound’s May 2026 Branded Link Update analysis, chat surfaces favor company-level mentions over flowery prose. Keep voice in the lower fold, where humans linger and bots have already moved on.
What canonical fields should every listing expose for a bot to shortlist it?
Seven fields carry the load: canonical title, brand string, currency value, availability enum, SKU, delivery window, and an image URL. Each must render as text inside the initial response. Per Similarweb’s May 2026 referral measurement, chat traffic lands on homepages 60 percent of the time. The listing has one fetch to prove it belongs on the shortlist.
Why do bots skip listings that hide the dollar value behind a quote form?
Because the bot has no hands to fill the form. It parses the initial fetch, scores it against the shopper’s brief, and moves on. A quote wall returns an empty value field, which fails the shortlist filter automatically.
Concept research from May 2026 on agentic search frames this as the second-reader problem. Visible numbers earn shortlist slots, hidden numbers earn silence.
When should catalog entries get attention before the homepage hero?
Catalog entries deserve attention once the homepage already names the brand cleanly. Per Profound’s May 2026 sector measurement, B2B software brands lifted over 200 percent above baseline while retail stayed flat.
The chat layer recommended the company, and the bot handled the shortlist later. Fix the front door first, then fix the listings the shortlist bot will fetch next.
How would you rewrite one product page first?
Pick one product page that should be earning agent-driven shortlist visits and is not. Open it in a private window. Save the rendered HTML to disk.
Now read only the HTML and ask the four questions in order. Can you find the price as plain text? The variant grid as plain HTML? The shipping ETA and the stock state as plain text?
If the answer to any of these is no, that is the rewrite.
Fix the structural fields first. Move the price out of the script, and mark up the product with the standard fields.
Then read the copy itself against the agent’s reading pattern. Rewrite the hero as a definition, move the proof block under the hero, and trim the brand vignette.
If your product pages live in WooCommerce or another store builder and you cannot tell whether the agent can read the prices, you can contact me here. Send the URL of one product page you would like to test, and the buyer question you would like the agent to ask. I will run the read-as-an-agent pass and name the one field that is currently invisible to the second reader.
No charge and no follow-up call. The product page that ships after the rewrite is the one the agent can put on the buyer’s shortlist.