Search is splitting into two channels. The classic result list still exists and still drives most commercial traffic. Alongside it, AI systems read the web and answer directly, and a growing share of research and comparison queries — which is where B2B buying decisions start — never produce a click at all.
Being the source that gets cited is becoming as valuable as ranking, and the optimisation for it is related but not identical. This is what we have found actually matters.
What changed about distribution
For twenty years the deal was straightforward: rank well, get the click, own the conversation. Answer engines change the shape of that. The system reads several sources, synthesises, and presents an answer with citations. You may influence a buying decision without ever receiving a visit.
That is not necessarily bad — being cited as the authority in an answer a decision-maker reads is valuable, and for B2B the eventual visit tends to be higher-intent. But it means optimising for extraction and citation rather than only for click-through.
It also raises the cost of vagueness. A page of marketing adjectives offers a summarising system nothing to quote, so it is passed over in favour of a page with specific, attributable facts.
How an answer engine actually reads your site
Three things matter most, in our testing and in what practitioners consistently report.
First, content must be in the initial HTML. Systems reading the web at scale frequently do not execute JavaScript, and even where they do, client-rendered content is read later and less reliably. A page whose body appears only after hydration may eventually be indexed; it will rarely be cited.
Second, structure must be semantic and descriptive. These systems extract sections rather than whole pages, so a heading that describes what the section actually answers — "How much does an ERP implementation cost in India" rather than "Pricing" — dramatically improves the chance of that section being selected.
Third, entity relationships must be explicit. Structured data is how a machine knows your organisation is the same entity across pages, that a service belongs to it, that an author works for it, and that an FAQ answers a specific question.
The entity graph, not the checklist
Most sites implement schema markup badly: a copy-pasted Organization block, an Article type on a service page, and FAQPage markup on content that is not actually a FAQ. Search engines increasingly ignore or penalise mismatched markup, so this is worse than nothing.
Think in terms of one connected graph rather than a list of tags. An Organization with a stable @id, referenced by every Service, Article, FAQPage and BreadcrumbList on the site. A LocalBusiness with complete address, geo-coordinates, hours and area served. Services with areaServed and a provider reference back to the organisation. Articles with an author who is a Person who worksFor the Organization.
And generate it programmatically from the same typed content that renders the page, so markup can never drift from what a human reads. Hand-written JSON-LD blocks decay the moment someone edits the page copy.
In practice
Content shapes that get extracted
Certain structures are reliably easier for a machine to extract, and they happen to be the structures human readers find most useful — which is why optimising for them is durable rather than a trick.
Direct question-and-answer pairs, with the question phrased the way people actually ask it. Comparison tables with real values rather than tick marks. Definitional passages that state what something is in one sentence before elaborating. Numbered procedures with a stated outcome. And specific, attributable facts: a number, a method, a named constraint, a measured result.
The test we apply to every important page: if a machine read this, what specific attributable fact could it quote? "We deliver world-class solutions" offers nothing. "A mid-sized warehouse migration takes four to seven months, of which six to fourteen weeks is dual running" offers something citable.
| Shape | Why it extracts well | Human benefit |
|---|---|---|
| Q&A pairs with natural phrasing | Maps directly onto a query | Answers the actual question fast |
| Comparison tables with values | Structured, unambiguous | Enables an actual decision |
| One-sentence definitions | Self-contained and quotable | Orients the reader immediately |
| Numbered procedures | Sequence is explicit | Followable without interpretation |
| Specific numbers with context | Attributable and checkable | Credible rather than vague |
What to stop doing
Client-rendering your main content. If the body requires JavaScript, you are gambling on crawler behaviour you do not control.
Marking up content you do not have. FAQPage schema on a page with no FAQ, Review markup you generated, Product markup on a service page. These are detected and they cost more than they gain.
Writing for word count. Long pages rank because depth serves the reader, not because length is a signal. A padded page offers a summariser more surface area for the same amount of substance, which does not help.
Keyword repetition. Answer engines work on meaning rather than string matching, and a page that repeats a phrase unnaturally reads as low quality to both machines and people.
Measuring something that resists measurement
This is the genuinely unsatisfying part and we say so rather than pretending otherwise. Classic search gives you impressions, positions and clicks. AI citation gives you very little that is directly observable, and the platforms expose different amounts of it.
What we do: track branded and long-tail query volume as a proxy, since being cited tends to raise branded search; monitor referral traffic from AI platforms where it is identifiable; periodically test a defined set of buying-stage questions against major answer engines and record whether the client appears; and watch for the pattern where enquiries arrive already informed, mentioning specifics from pages nobody clicked.
It is imperfect and improving. What is not in doubt is that the underlying work — server-rendered content, clean semantic structure, correct entity markup, specific citable facts — improves classic search performance too. So the investment is defensible even before the measurement matures.
Key takeaways
- Content must be in the initial HTML — client-rendered bodies are rarely cited.
- Headings should describe what a section answers, because engines extract sections rather than pages.
- Build one connected entity graph with a stable Organization @id, generated from typed content.
- Favour Q&A pairs, comparison tables, definitions, procedures and specific attributable facts.
- Measurement is immature, but the underlying work improves classic search performance regardless.
Frequently asked
Overlapping rather than separate. Server-rendered content, clean semantic structure, correct entity markup and genuinely useful writing serve both. The differences are emphasis: answer engines reward specific citable facts and clear section structure more heavily, and reward link authority somewhat less. If you are doing technical SEO well, you are most of the way there.
No. It makes you eligible and legible; it does not guarantee anything. Markup that does not match visible content is worse than none, because mismatch is detected and penalised. Treat it as making your content readable to machines rather than as a lever to pull.
A genuine strategic decision rather than a technical one. Blocking protects content from being summarised without a visit; it also removes you from answers your buyers are reading. For most B2B businesses, being cited is worth more than the click you might have received. For publishers whose revenue is impressions, the calculation is different. We will not make this call for you, but we will make sure it is a decision rather than an accident.
Technical fixes — server rendering, structured data, section structure — show in classic search within four to eight weeks. Citation behaviour is slower and less observable, and depends on your content being genuinely the best available answer for a specific question. This is not a shortcut; it is the same work as being genuinely useful, done deliberately.