PART TWO — THE FIVE THINGS THAT DECIDE IT
Chapter 8 — Be Specific
Answer engines invert the old rule of going broad first. Because a question is broken into sub-questions and matched on proximity rather than aggregate reputation, the page that addresses one exact intersection wins retrieval for it. This chapter covers specificity, the interventions with measured lift, and the order to build in.
From Becoming the Answer by Jeremy Osborn · 1,082 words
Traditional search rewarded going broad first. Build category pages for high-volume terms, accumulate authority, let it lift everything beneath.
Answer engines invert this, for a mechanical reason rather than a philosophical one.
When someone asks a specific question, the system breaks it into sub-questions targeting each attribute. A company with content addressing exactly that intersection wins retrieval for it, largely regardless of overall authority. In the vector space where this happens, you compete on proximity to a specific question, not on aggregate reputation.
That’s the opening for specialists, and it’s real. A niche company with authoritative, data-rich content on a narrow domain sits closer to a specific question than a generalist whose coverage is broad and thin. Large companies’ legacy content libraries — optimized for high-volume head terms — often produce few passages that survive chunking and comparison against a precise attribute combination.
But hold that against Chapter 4. Roughly 70 percent of AI Overview citations come from domains already ranking on page one. Specificity is the differentiator within the retrievable set. It is not a substitute for being in it.
The honest formulation: authority gets you into the room. Specificity wins the argument inside it. You need both, and most organizations have exactly one.
Building a question inventory
This is the core new workflow. Two weeks the first time, a day per quarter after.
Step one: harvest real language. Not invented language. In rough order of value:
Sales call recordings — the questions prospects actually ask, in their own words. This is the richest source in most organizations and nobody in marketing has ever listened to it. Then support tickets and chat logs. Your site’s internal search. Sales objection logs, which are comparison questions in disguise. Community threads in your category. Search Console queries, still useful, now one input among several.
Step two: build the attribute matrix. For each core buying question, list the constraints a real buyer applies — company size, industry, integrations, budget, deployment model, compliance regime, sophistication, the specific job to be done. Then generate the intersections.
Not “best project management software,” but:
project management software for a 15-person construction firm that needs offline access
project management tool that works with QuickBooks and doesn’t charge per seat
simplest project management software for a team that has never used one
You aren’t writing a page for each. You’re mapping the sub-question space you’re being evaluated against.
Step three: classify.
| Type | Example | Priority |
|---|---|---|
| Category definition | “what is X” | Low — you’ll lose to Wikipedia |
| Comparison | “X vs Y” | High — commercial, and produces more brand mentions |
| Constrained recommendation | “best X for [situation]” | Highest — the sweet spot |
| Implementation | “how do I do X with Y” | High — proves experience |
| Troubleshooting | “why does X do Y” | Medium — strong experience signal |
| Brand-specific | “is [you] good for X” | High — reputation defense |
Step four: score for winnability. Three checks per high-priority question. Who currently gets cited — run it several times and record the domains. Do you rank on page one for the closest search equivalent; if not, that’s your first job. And: do you have information nobody else has for this question?
That third test kills more content plans than any other, and it should.
Information gain
Chapter 4 was hard on content tactics. Formatting effects measured near zero. Rewrites sometimes hurt. But one thing survived in every study that found anything at all: content carrying information a competing passage cannot.
In the academic work on this, the three highest-performing interventions were adding quotations from recognized authorities (+41 percent), adding statistics (+31 percent), and citing sources (+27 percent). Keyword stuffing went backwards, at −8 percent — the worst-performing method tested. In the largest controlled experiment, the presence of real price information and evidence-backed claims carried some of the largest effects measured.
The mechanism is the head-to-head comparison from Chapter 3. Faced with two passages, the system prefers the one offering something the other doesn’t. A verifiable number. A named source. A real price. A dated measurement. A documented case. A passage that restates common knowledge in slightly different words offers nothing to prefer.
In practice:
Publish real numbers. Your data, your benchmarks, your pricing. If you won’t publish price, you’re absent from one of the strongest content factors anyone has measured.
Name your sources. Attribution is a retrieval asset, not an academic courtesy.
Date everything. An undated page is an old page.
Publish specifications. Missing specs measured as a significant negative. Tables of real attributes beat prose about benefits.
Write from experience. Case studies, first-hand accounts, practitioner data. As the web fills with generated content, evidence of actual involvement appreciates.
Be direct. Hedged writing measured worse than confident writing. That’s not license to overclaim. It’s license to state what you know plainly.
One caution: these advantages compress as competitors adopt them. Information gain isn’t a permanent moat. It’s a moat exactly as long as you actually have information others don’t — which is an argument for original research over content production.
The build order
Tier one: constrained recommendation pages. The exact intersections your best customers describe. Real specs, real prices, real constraints, and honest statements about who this isn’t for. Unglamorous, invisible in a keyword tool, and they win sub-questions.
Tier two: comparison content, including against competitors, written fairly enough that the system will use it. Fairness is instrumentally optimal here — a comparison that reads as marketing gets discounted during grounding, because it contradicts everything else the system found.
Tier three: implementation and troubleshooting depth. The material that proves experience, and the material people link to in forums.
Tier four: original research. The most durable asset, and the one most likely to be cited by the third-party sources from Chapter 7. One well-designed annual study beats fifty blog posts.
Tier five: category education. Last. You’ll lose “what is X” to Wikipedia, and it’s the traffic least likely to convert.
Three things not to do
Don’t mass-produce permutation pages. Generated combinations fail on information gain by construction, and content volume is the weakest correlate on the entire list. Fifty thin pages lose to five substantial ones and cost you crawl budget you’re already wasting.
Don’t restructure your library for “chunkability.” The evidence isn’t there.
Don’t write for the machine at the reader’s expense. The benchmark finding that conversational rewrites often hurt ranking is the empirical version of something that should already be obvious. These systems are trained on human preference. Writing that reads as machine-directed is a detectable signal, and detectable signals get discounted.