GEO-043·Issue 04·Technical Foundation·43 / 132
FAQ, HowTo, Review: Which Structured Data Still Works?
Google's rich results keep shrinking, but machine comprehension does not. Audit markup field by field, fix what is invisible, and know when to skip schema.
Issue focusAnswer Units: Writing Category and Product Pages for GEO
On a Thursday in April, a DTC brand selling portable power stations into the US and Germany reviews its rich-results report. FAQ and HowTo have been flat for two quarters; Review snippets survive on a few product pages. The marketplace lead asks why an engineer still maintains the JSON-LD. The tech lead asks a harder question: when a model summarizes a PDP, does it know which line is the question, which is the answer, and which rating is real?
Structured data’s first job is to make the page unambiguous to a machine; the rich result is a platform-controlled bonus, so fund the markup for comprehension and treat display as upside.
Rich results are rented; the markup is the asset
Annotation turns strings into typed facts: a question, an answer, a rating and its count. Display is the engine’s choice; the first layer is yours, the second is not.
As of 2026, Google limits FAQ rich results to a small set of authoritative sites, no longer lists HowTo rich results, and keeps review snippets for some content types under provenance rules.
Comprehension outlives the snippet. An extraction pipeline still has to separate spec from marketing and opinion from claim; clean Q&A boundaries, explicit step order, and a rating tied to a visible count reduce that ambiguity. Markup describing invisible content adds confident facts no one can check.
Ask not “does this earn a rich result?” but “if the snippet never returns, is this worth maintaining?” If no, delete it.
Three types, three different risks
Sold as one package, they behave differently.
| Type | What it still earns | Where the value sits | What goes wrong |
|---|---|---|---|
| FAQPage | Rich results heavily restricted | Q&A separable; the answer lifts whole | Questions absent from the page |
| HowTo | No rich result | Steps, tools, time, cautions made explicit | Checkout written as a procedure |
| Review / AggregateRating | Snippets for some content types | Rating and count become comparable fields | Self-serving or mismatched numbers |
FAQPage’s display value has been cut back hard; its remaining value is structural. A clean answer can be extracted whole, as the answer unit framework describes — if the question exists on the page. Hidden Q&As built for search volume read as keyword filling.
HowTo lost its rich result, but “how do I use this” is real in every physical category: first charge, filter replacement, wall mounting, cleaning. Explicit steps and cautions give a model something to repeat. The counterexample is checkout written as a HowTo: a transaction flow is not a procedure, and passing validation does not fix the semantics.
Review still affects display in some contexts and carries the most risk. Ratings become comparable fields, so mismatches show: 27 reviews on the page, 132 in the markup. That reads as intent to mislead, which is what snippet policies target. And if Product and Offer are wrong, FAQ polish buys little — see Product schema in 2026.
Reconcile first, then assign, then generate
Order matters more than the checklist. Most teams copy the newest template before checking it.
| Step | Action | Owner | Passes when |
|---|---|---|---|
| 1 Reconcile | Compare every markup field with the rendered page | Engineering and content | Each field has a visible source, or is deleted |
| 2 Assign | One objective per type: be understood, or earn display | Growth | One stated objective per type |
| 3 Generate | Emit markup from templates and product master data | Engineering | A new SKU needs no manual edit |
| 4 Verify | Validator, then prompt replay for extraction quality | Engineering and growth | Zero errors; consistent fields on 20 pages |
Step 1 is easy and usually skipped. Put markup beside the page and problems surface: answers living only in JSON, a price from three months ago, a review count never synced. A field with no visible source is deleted.
Steps 2 and 3 are the discipline. Write one objective per type — extraction quality or display — then generate the markup from the template and product master data. Hand-written JSON-LD drifts within a quarter; template output survives a price change.
Step 4 has two halves: the validator proves syntax, a prompt replay proves the right sentences were taken. A clean validator with a wrong extracted answer is a false positive; test the replay against the AI visibility baseline.
Example: deleting half the FAQ block
Example — a hypothetical power-station brand, no real company or numbers. Its PDP template emits 18 FAQs, 11 of which never render. AggregateRating claims 4.9 from 132 reviews while the platform holds 27. A HowTo explains how to order.
Delete the 11 invisible Q&As; keep the seven that exist, lift answers into visible text, and cut each to two sentences. Reset the rating to the real average and count, and show the star distribution. Replace the checkout HowTo with a visible block on first charge and running two units in parallel, listing tools, time, and cautions.
Acceptance is not a returning rich result. It is three checks: every field maps to visible content, the validator is clean, and ten “how do I use it” prompts replay with the right steps and limits.
Where this breaks
First, when revenue concentrates in marketplaces. On Amazon, Shopee, or TikTok Shop, the assistant reads listings, reviews, stock, and fulfillment; JSON-LD has near-zero influence. Fix the feed first.
Second, when there are no real reviews. A zero-review SKU should not carry an AggregateRating to look established — a policy problem before a marketing one. Use test reports, certifications, and warranty terms, as in first-party evidence.
Third, when the content has no answer. “We are committed to quality” marked up correctly is still an empty sentence, and extracting it damages more than silence.
Fourth, when the site already has a structured-data manual action. Clean the inventory before adding anything; new markup during a problem period extends recovery.
Fifth, when schema is treated as a ranking switch. No official documentation promises higher rankings or entry into an AI answer. It lowers the chance of being misread and raises the chance of being quoted correctly — nothing more.
This month, take the PDP, category, and help-center templates. Export every field they emit, set each beside the page, delete what you cannot point to, and move what survives into template output with a named owner. Then debate adding another type; the technical foundation hub indexes this line.