
Ranking optimises for a page being chosen. Extraction optimises for a passage being liftable. These are different objectives, they have different failure modes, and the second one is now doing a lot of the work that the first used to.
The distinction matters practically, because a page can rank respectably and be functionally invisible to an assistant, and the fixes for each problem barely overlap.
Why a good page can be unliftable
Extraction operates on passages, not documents. A summariser needs a span of text that answers a specific question and survives being removed from its surroundings.
Most well-written marketing copy fails that test for structural reasons that have nothing to do with quality.
It builds context before arriving at the point, so the first two sentences under a heading are preamble. It uses pronouns and back-references — "this approach", "as mentioned above" — which break the moment the passage is lifted. It distributes a single answer across three paragraphs, so no contiguous span contains the whole thing. And it hedges, because hedging reads as professional, but a qualifier without a condition carries no extractable information.
Every one of those is a habit that makes prose read better and makes it unliftable.
The passage contract
The structural rules that follow are mechanical enough to check.
One question per heading, phrased as a customer would ask it. Not "Our Pricing Philosophy" but "How much does a boiler replacement cost?" The heading is the retrieval key; writing it in your own vocabulary rather than the customer's is the most common single error.
The answer in the first one or two sentences. Then elaborate. This inverts normal essay structure and feels abrupt to write, and it is the single highest-leverage change available.
Self-contained passages. No pronouns pointing outside the section, no "as we discussed", no dependence on a table three headings up. If it cannot stand alone, it will not be lifted, and if it is lifted anyway it will be lifted wrong.
Concrete values with units and named conditions. "Typically eight to sixteen weeks, assuming content is ready at kickoff" is extractable. "Timelines vary depending on project complexity" is not, because it contains no information.
Tables for comparisons. Structured comparisons survive summarisation intact. Prose describing the same comparison gets flattened into something approximately true.
Content modelling is the real fix
Enforcing the above by convention across a growing site does not work. Enforcing it by data model does.
Model a service as a structured entity rather than as a hand-authored page: name, plain-language description, price or price range, duration, service area, prerequisites, and a list of question-answer pairs. Then derive the visible page, the structured data, the booking flow and any comparison table from that single definition.
Three things follow, and all of them matter.
The markup cannot disagree with the visible content, because both are projections of the same source. Markup that contradicts what a human sees is a durable liability, and this makes the contradiction structurally impossible rather than merely discouraged.
The passage contract becomes a schema constraint. If the model requires a short plain-language summary field, every service has one, and no writer can quietly omit it under deadline.
And adding a new output — a comparison page, a printable quote sheet, a machine-readable feed — is a template, not a content migration.
Structured data as a graph, not fragments
Related point at the markup level.
Emit one connected @graph with @id references: the business node, service nodes referencing it as provider, and FAQPage entities attached to the pages that actually contain visible question-and-answer content.
Scattered independent JSON-LD blocks are individually valid and collectively uninformative, because nothing expresses the relationship between a service and the business providing it. The relationships are the point.
For the business node, include openingHoursSpecification with real holiday exceptions, areaServed, sameAs for profiles the business controls, and pricing on services where the business is willing to publish it. Price is the field most often omitted and most often asked about.
The thing that stopped working
Worth stating plainly because the tactic is still being sold: generating many near-identical pages by templating a variable — town names being the classic — is now a reliable way to have the whole set discounted rather than a way to expand coverage.
The extraction-era logic is straightforward. If thirty pages contain the same passage with one noun changed, none of them is a distinctive source of anything. One page per location genuinely served, containing real local specifics, both performs better and is less work.
Performance as a precondition
One non-content prerequisite. Assistant-referred visitors arrive transactional and impatient, and mobile abandonment rises sharply between two and five seconds.
The recurring offenders are predictable and each is roughly a day: uncompressed hero images without explicit dimensions, render-blocking third-party scripts loaded before content, web fonts without font-display: swap, and map or video embeds loading full payloads at page load rather than on interaction.
Put a target in the contract, measured on a mid-range phone over a throttled connection, and make final payment contingent on it.
Buyer-side guide with pricing bands and vendor checklist: Web Development Services for Small Business: A 2026 Buyer's Guide. For requirements that have outgrown a site, see web app development.
Frequently Asked Questions
What is the difference between optimising for ranking and for extraction?
Ranking optimises for a page being selected; extraction optimises for a passage being liftable and self-contained. A page can rank respectably while containing no span an assistant can quote.
Why does well-written marketing copy often fail extraction?
Because it builds context before the point, uses back-references that break when lifted, spreads one answer across several paragraphs, and hedges without naming conditions. These habits improve readability and prevent extraction.
How should a heading be phrased for extraction?
As the customer would ask the question, not in internal vocabulary. "How much does a boiler replacement cost?" rather than "Our Pricing Philosophy". The heading functions as the retrieval key.
Why model services as data rather than pages?
Because one definition can derive the page, the structured data, the booking flow and comparison tables — making markup and visible content structurally unable to disagree, and turning the passage contract into a schema constraint.
Why emit a single connected @graph?
Because relationships carry the meaning. Independent JSON-LD blocks are valid but express nothing about which business provides which service; @id references make the entity graph legible.