Schema markup for AI answer engines
what still earns its keep

Does schema markup help with AI answers?

It depends entirely on whose AI you mean, and the two largest platforms are on record saying opposite things.

Google, in its guide to optimizing for generative AI features published 15 May 2026: structured data is not required for generative AI search, and there is no special schema.org markup you need to add. That is unambiguous and it comes from the platform itself, not from an inference.

Microsoft, through Fabrice Canel, principal product manager at Bing, speaking at SMX Munich in March 2025: schema markup helps Microsoft's language models understand content, and site owners should push updates through IndexNow. That is the single strongest platform level justification for schema in the AI context available anywhere, and it is worth noticing that it does not come from Google.

Neither statement contradicts the other, because Google is describing an input to its own generative features and Microsoft is describing how its models parse content. Perplexity and Anthropic have published nothing on the subject either way, so anyone telling you Perplexity weights schema is inferring, and should say so.

So what is schema actually for now? Three things, in descending order of value. It disambiguates entities, which is the thing that determines whether an assistant knows your company is your company and not a similarly named one. It remains an eligibility input for whatever rich results still exist in classic search. And it is a machine readable statement of relationships that costs almost nothing to maintain once it is built. What it is not, in 2026, is a lever you pull to increase citations.

Is FAQ schema still worth adding?

Not for a Google rich result, because that feature is fully deprecated. Google restricted FAQ rich results to authoritative government and health sites in August 2023, then removed the feature entirely: the deprecation notice was added to Google's documentation on 8 May 2026, the feature stopped appearing on 7 May 2026, the FAQ search appearance filter and rich result reporting were removed in June 2026, and Search Console API access wound down by August 2026.

Google's documentation notes the markup can stay in place and causes no problem. It simply produces nothing visible.

Here is the distinction that matters and that most coverage of this deprecation missed. The FAQ section as a content structure is strongly supported and worth keeping on every substantial page, because a question followed by a direct answer mirrors the shape of the query a retrieval system is matching against. The FAQPage markup is now a parsing convenience with zero Google upside. Keep the section. Ship the markup if it costs you nothing. Do not let anyone report it as a win, and be suspicious of any 2026 proposal that leads with FAQ schema as a GEO tactic, because it tells you when the author last checked.

Which schema types are worth shipping in 2026?

Six worth shipping, two worth skipping, one worth an honest shrug.

How do you make a language model resolve your brand correctly?

By saying the same thing about yourself, in the same words, in as many machine readable places as you control, and then getting sources you do not control to agree.

Language models resolve entities by corroboration. They are not looking up a record, they are reconciling many statements. When your LinkedIn page says one company name, your schema says a second spelling, your press coverage says a third, and your product docs use a fourth, you have not given the model a brand, you have given it a disambiguation problem. It will resolve that problem in whichever direction the weight of evidence points, and it may not be your preferred direction.

Three practical rules come out of that.

Pick one display name and one alternate

Ship the canonical name as name and the variant as alternateName in your Organization block. That single field is what tells a machine that two strings are one entity. We run "Outbound Pros" as the display name with "OutboundPros" as the alternate on every property in the group, because both strings are already live across the estate and pretending otherwise would fragment the entity rather than fix it.

Give every author one identifier and reuse it byte for byte

Our founder's Person node uses the same identifier on all five group properties, which is what collapses five author references into one author entity. A mismatched identifier across properties produces five strangers who happen to share a name, which is worse than having no Person markup at all. Worth noting that an identifier is not a link: it is a string that has to match exactly, and pointing it at a page that does not serve real content yet would be worse than leaving the link off, which is why the author name on this page is plain text.

State corporate relationships explicitly

What schema will get you in trouble?

Fabricated review markup, and it is the one schema mistake with legal exposure attached rather than just a ranking cost.

Rating and review blocks describing testimonials that no named customer has approved in writing are a documented Google structured data manual action trigger. That is the cheap version of the consequence. The expensive version is the United States Federal Trade Commission's Rule on the Use of Consumer Reviews and Testimonials, 16 CFR Part 465, published in the Federal Register on 22 August 2024 and in force since 21 October 2024. Penalties are civil and assessed per violation, the rule has not been vacated, and the FTC issued warning letters under it on 22 December 2025, so it is being enforced and not sitting on a shelf. We publish no penalty figure here, because the commonly circulated number is a prior year adjustment and we have not verified the current one. The EU Omnibus Directive requires the same disclosure of connected reviews.

Two sections of that rule govern how anyone should think about testimonial content on their own site. Section 465.5 prohibits a business disseminating testimonials by employees, officers, managers or agents without disclosure when the relationship is not otherwise clear to the audience. Section 465.6 makes it an unfair or deceptive practice to materially misrepresent, expressly or by implication, that a website or entity the business controls provides independent reviews or opinions about a category of businesses including its own.

Read section 465.6 carefully, because the line it draws is narrower and more workable than most people assume. The prohibited act is misrepresenting independence, expressly or by implication. Owning the site is fine. Publishing opinions on it is fine. Recommending your own company on it is fine. Letting a reader believe the site is an independent evaluator when it is not is the violation. Note the phrase "or by implication": silence about ownership on a site that reads like an independent comparison resource is itself the implication.

What do we not know about schema and AI?

We do not know whether any answer engine consumes Dataset, SoftwareApplication, or most of the long tail of schema.org types. There is no evidence either way, no platform has commented, and we ship them because they are cheap and descriptive rather than because we have seen them work.

We do not know whether Perplexity uses schema markup at all. Perplexity has published crawler documentation and nothing on parsing. Every confident answer you will find on this question is inference presented as fact.

The honest summary is that schema in 2026 is a hygiene and identity layer with one platform endorsement, one platform denial, and silence from the rest. Ship the six types above, spend a day on your entity graph, and put the rest of your effort into the thing that is actually confirmed to matter, which is having something worth quoting on the page.