Answer Engines Read Tables, Not Taglines
A prose hero section tells a human buyer why your product feels modern. ChatGPT, Claude, and Perplexity skip the adjectives and cite whichever competitor published a pricing table and a plain spec sheet.
Ask ChatGPT, Claude, or Perplexity for the best Postgres host or issue tracker for a 40-engineer team, and watch which URLs actually get cited.
Across 1,800 buyer-intent pulls on Stratos (600 prompts × 3 engines), better-funded incumbents routinely lost the #1 recommendation to smaller startups for a remarkably mundane reason: the incumbent hid limits behind 'Contact Sales' prose, while the startup published a plain HTML table.
Why do answer engines skip polished marketing copy?
Retrieval-augmented answer engines work on tight token budgets. When a user asks a comparative buyer question — latency ceiling, SOC 2 tier, per-seat overage — the retrieval pass scores passages with high factual density.
Adjectives have zero factual density. A paragraph promising 'effortless enterprise scale' gives the synthesizer nothing it can safely quote without guessing, so the model pulls numbers from a competitor's comparison doc or a third-party GitHub benchmark instead.
Why does a single prompt screenshot lie about rank?
The second mistake teams make is running one prompt in an incognito tab, screenshotting position #1, and calling it a growth metric. Live web search fan-out is stochastic.
Running the exact same prompt sixteen times within an hour routinely flips ranks #1 and #3 on individual pulls. Only when you sample in batches (N >= 8) and extract structured citations in a separate pass do real Share-of-Voice shifts separate from retrieval noise.
| Page Surface | ChatGPT | Perplexity | Claude |
|---|---|---|---|
| Hero & narrative landing copy | 6% | 4% | 9% |
| Plain pricing + limits HTML table | 38% | 44% | 34% |
| Public X-vs-Y docs with exact caps | 31% | 29% | 35% |
| Third-party benchmarks / Reddit | 25% | 23% | 22% |
{
"prompt_cluster": "postgres-branching-50-seats",
"sample_runs": 12,
"generation_mode": "natural_prose_with_web_search",
"extracted_rank_1": {
"brand": "Neon",
"share_of_voice": 0.46,
"confidence_band": "±0.05",
"cited_surface": "/docs/introduction/plans#compute-limits-table"
}
}Spec tables lift factual density; answer engines quote exact caps and pricing.
↑ Switch page structure (prose vs. spec tables) and drag prompt sample count (N) to see how single-shot rank flips settle into real Share-of-Voice.
What structured tables still won't fix
- Publishing comparison tables won't rescue a product with negative Reddit or Hacker News post-mortems — Perplexity and ChatGPT weight independent complaint threads heavily.
- Forcing strict JSON mode during the initial LLM query warps live search tool-calling; structured extraction has to run as a second pass over natural prose.
- Pre-cutoff model weights change slowly for unsearched prompts; table changes move live-search answers in days, not pure offline completions.
Operating Rules
- Replace narrative feature walls with explicit HTML tables stating hard limits, regions, and unit prices.
- Never treat N = 1 prompt checks as ground truth; measure stratified batches with dispersion bands.
- Separate natural user-mode sampling from schema-constrained citation extraction.