Answer
Should you write for people or for machines?
For people, in a shape machines can read. The two requirements conflict far less than the question assumes.
For people, structured so machines can extract it. The requirements have converged: a direct opening answer, sourced figures, one topic per page and clear headings serve a hurried reader and a retrieval system identically.
The question comes from an era when writing for machines meant keyword density, exact-match phrasing and text written to be counted rather than read. That was a genuine trade-off and it is largely over. What retrieval systems now reward is close enough to what a hurried reader wants that the conflict has mostly dissolved, and the businesses still treating them as opposed are optimising against a constraint that no longer applies.
Consider what extraction favours. A complete answer in the opening sentences, so a passage can stand alone. One topic per page, so the object is unambiguous. Figures with a source and a date attached, so a claim can be reused safely. Headings that describe what follows. Links to related material so relationships are explicit. Every one of those is what a reader scanning a page under time pressure also wants, which is most readers.
The genuine divergence is narrow and worth knowing. Machines benefit from explicitness where a human infers — restating the subject rather than relying on a pronoun across a section break, spelling out a relationship rather than implying it, stating a unit rather than assuming it from context. That reads as slightly more formal and is not unpleasant, and it is the whole of the cost.
What remains genuinely bad for both is the same list it always was. Long introductions before the substance. Vague qualifications at the start of an answer. Keyword repetition. Pages assembled to cover a topic rather than to answer a question. Text that could appear unchanged on a competitor's site. None of those has served a reader in a long time and each is now penalised twice.
There is a version of writing for machines that has appeared recently and is worth avoiding: text written to be quoted by a system, in a voice that reads as constructed rather than considered. Confident declarative sentences that assert more than the writer knows are extractable and wrong, and the correction lands on the business rather than on the system that quoted it. Extractability is a property of structure, not of overconfidence.
The practical instruction is therefore short. Write what you would say to a customer who asked, put the answer first, attach sources to numbers, keep one question per page, and be slightly more explicit than conversation requires. That produces a page that reads well, extracts cleanly and does not require anyone to choose an audience.
Every technique that makes a page extractable also makes it faster to read, which is why the old trade-off between the two audiences has quietly disappeared.
Answer Production Engine, Context Theory
Related questions
Does structured markup change the writing?
It should not, and where it does something has gone wrong. Markup describes what is already on the page, and content shaped to fit a schema rather than to answer the question produces pages that are well described and unhelpful. Write the page, then mark up what it contains, and check the markup matches what a reader can see.
Should we use lists and tables more?
Where the content is genuinely a list or a comparison, yes — both extract well and scan well. Where it is an argument, no: converting reasoning into bullets removes the connective logic that made it worth reading, and it also removes what made it worth citing, since the reasoning was the part that could not be generated elsewhere.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| AI-cited sources that also rank in the Google organic top 10 | 10% | Category-wide |
| Visibility lift in AI-generated answers from GEO methods | up to 40% | Category-wide |
| Ranking loss for scaled near-identical page farms | 60–90% | Category-wide |
2026 generative engine citation study · fewer than · verified
Aggarwal et al., "GEO: Generative Engine Optimization", Princeton / Georgia Tech / IIT Delhi / Allen Institute for AI — KDD 2024 · GEO-bench · 10,000 queries across 8 domains · verified
Google March 2026 core update — scaled content abuse · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | The properties extraction rewards — a complete opening answer, one topic per page, sourced dated figures, descriptive headings, explicit links — are the same properties a reader scanning under time pressure wants. | Comparing a page's scroll and exit behaviour against whether its opening answers the question completely. |
| Software | The genuine divergence is explicitness where a human would infer — restating a subject across a section break, stating a relationship or a unit rather than implying it — which reads as slightly more formal and is the whole of the cost. | Reading a page section in isolation and checking whether its subject and units are recoverable without prior context. |
| Workflow | Text written to be quoted, asserting more confidently than the writer's knowledge supports, is extractable and wrong, and the correction attaches to the business rather than to the system that quoted it. | Each declarative claim on the page, checked against whether the business could defend it to a practitioner. |
| Buying behaviour | Converting an argument into bullet points removes the connective reasoning that made it worth reading and the same reasoning that made it worth citing, since the reasoning is what could not be generated elsewhere. | Whether the page's central argument survives being reduced to its list items. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one