Writing product copy that survives being summarised

Whatever the system — an AI Overview, an assistant answering a shopping question, a sales rep pasting your page into a chat window — the mechanism is broadly the same. Something retrieves a chunk of your page, decides it is relevant to a question, and composes an answer partly by paraphrasing it and partly by reproducing sentences from it more or less intact.

Two consequences follow, and they are the whole of this article. First, the sentence is the unit of survival: a claim that needs the sentence before it in order to make sense is a claim that arrives broken or not at all. Second, compression is lossy in a predictable direction — it keeps the assertion and drops the qualification, because the assertion is what answers the question.

None of this is a new discipline. It is the same skill a journalist uses writing a paragraph that will be cut from the bottom, or a lawyer writing a clause that must survive being quoted. It happens to be the skill product copy is least often written with.

It is also worth saying what is not required, because a lot is being sold on the premise that something is. Google’s documentation on AI features states plainly that there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimisations necessary, and that the bar is a page being indexed and eligible to be shown in Search with a snippet. The work is in the prose, not in a file in your root directory.

The single highest-value habit, and the one that feels most unnatural to anyone trained to write elegantly. Pronouns are cohesion devices: they bind a sentence to the one before it. That binding is exactly what breaks when a sentence is extracted.

“It holds 32 litres and fits under most airline seats” is a good sentence in a paragraph and a useless one on its own. “The Harlow 32 holds 32 litres and fits under most airline seats” is the same sentence with the antecedent restored, and it is portable.

Applied literally to every sentence this produces copy that reads like a robot wrote it, so apply it selectively: restate the subject in the sentences that carry a fact somebody might ask about, and let the connective tissue between them use pronouns freely. A practical test is to read the page and ask, of each sentence, whether a stranger could be shown it alone and told what product it describes. If not, and the sentence contains a fact worth having, fix it.

The same rule applies to headings. A fact that exists only in a heading — “Waterproof” above a paragraph about the lining — is a fact that a passage may arrive without. Repeat it in the body text underneath in a full sentence. Redundancy that looks clumsy in a document is what makes a fragment survive.

This is the failure with real consequences, because it produces summaries that are wrong about your product in a way that reflects on you rather than on the system that got it wrong.

Write “Water resistant to 50 metres.” in one sentence and “The crown must be fully screwed down.” in the next, and you have invited a summary that says the watch is water resistant to 50 metres. That statement is now circulating without its condition, and the first customer to swim with an unscrewed crown will take it up with your support team.

The fix is to refuse to separate them: “Water resistant to 50 metres with the crown fully screwed down.” Longer, less elegant, and it cannot be truncated into a false claim without also losing the claim.

The same applies to every qualifier a product page routinely strands in a neighbouring sentence — “on orders over £50”, “excluding the Highlands and Islands”, “when fitted to models from 2019 onwards”, “sold separately”. Each of those is a condition on a promise. If it is not in the same sentence as the promise, assume it will not arrive with it.

Numbers deserve the same treatment. Put the unit adjacent to the figure, spell out what the figure measures, and avoid comparatives that depend on a referent stated elsewhere. “40% lighter” is meaningless in isolation; “40% lighter than the previous Harlow 32, at 780g” is a fact.

Structure helps retrieval — a table row, a list item and a short paragraph are all clean passage boundaries — but do not let structure be the only thing carrying meaning, because structure is the first thing removed in transit.

Shopify makes this concrete. The Admin API documents the product description field as a single-line description of the product with HTML tags removed, alongside descriptionHtml, which preserves the markup. Plenty of things read the stripped version: feeds, marketplace integrations, internal tools, anything that wanted plain text. A specification you expressed as six list items becomes six clauses running together with no separators, and “Material: cotton Weight: 780g Fits: 15in laptop” is worse than the paragraph you did not write.

The same principle applies to tables. A cell reading “50” is meaningless once separated from its column header, so put the unit in the cell: “50 m”. A row label of “Weight” with a cell of “780” is a table designed for a human eye scanning a grid, not for a fragment being lifted out of one.

And write list items as complete sentences rather than fragments hanging off the lead-in above them. A bullet reading “under airline seats” depends entirely on a stem sentence that will not come with it.

Google documents a small set of levers, and they are worth knowing precisely rather than approximately, because two of them are easy to over-apply.

data-nosnippet marks text that should not be used in a snippet. Google documents it as working on span, div and section elements only — an attribute on any other element does nothing — and warns that because the attribute may be extracted before or after rendering, you should not add or remove it with JavaScript on existing nodes. Use it for genuine boilerplate: shipping banners, legal furniture, the cookie notice. Do not use it on anything a customer might want answered.

max-snippet:[number] caps the characters used, with 0 equivalent to nosnippet and -1 letting Google choose the most effective length. Google notes the setting applies across result formats including AI Overviews. There is very rarely a good reason for a commerce site to cap this; the instinct to limit what gets shown is usually the instinct to keep people clicking, and it trades away the answer that would have earned the click.

The honest framing: these are controls for suppressing text, not for promoting it. There is no markup that makes a sentence more quotable. Writing the sentence properly is the mechanism.

Everything above is achievable on a page you are writing by hand this afternoon. Doing it across a catalogue is a different problem, and it is not solved by asking merchandisers to write better.

The parts that scale are the ones the template owns. A specification table rendered from structured fields, with units in the cells, is written once and correct on every product. A returns and delivery block that states the actual terms in full sentences is written once. A compatibility statement pattern — including the negative case — becomes a field somebody fills rather than a paragraph somebody composes.

What does not scale is any of this arriving through a developer ticket. The stores that publish good answer content are the ones where a merchandiser can build a comparison table, a numbered guide or an FAQ block on a Tuesday without a deploy, which is the argument for a composable [Shopify section library](/services/build/shopify-section-library) rather than a set of fixed page templates. The writing advice in this article is cheap and widely available. The ability to apply it repeatedly, without a sprint, is the thing most stores are actually short of.

Rewriting the whole catalogue. Start with the pages that already get traffic and the questions support already receives. A hundred products rewritten properly beats four thousand rewritten by a generator that adds fluency and no facts.

Writing for the machine at the expense of the reader. Every technique here — name the subject, keep the condition with the claim, state the limits — makes copy clearer for a person skim-reading on a phone. If a change makes the page worse to read, it is the wrong change; the overlap between the two audiences is nearly total, and where they diverge the human wins.

Treating an AI visibility score as a measurement. Sample a fixed set of real buyer questions across the assistants your customers use, record the answers and the sources, and repeat monthly by the same method. Crude, and it catches the thing that matters most: a false claim about your product, stripped of the condition you failed to attach to it, repeating across systems.

• Shopify section library
• Answer engine optimisation for a product catalogue
• Getting found by AI assistants