Only 3.6% of AI Search Patents About Ranking or Selecting Sources Mention Freshness, While 36.9% Mention Structured Data

Summary

Picture a café owner wondering what makes an AI assistant pick one café over another. Patent filings will not answer that for any specific product, but they do show which signals inventors write into their methods for choosing sources.

We took 358 generative AI search patent families that describe ranking or selecting sources and checked six groups of signals. Structured data and knowledge bases lead by far, in 132 families (36.9%). Freshness shows up in 13 (3.6%), and reviews in 9 (2.5%).

Most families, 191 of 358, name none of the six. A patent shows what was filed, not what is live in any product. These counts describe patent text, not how any company ranks anything.

What we measured

We pulled every publication in Google Patents Public Data (patents-public-data.patents.publications) published from January 1, 2018 to September 17, 2026 that carries at least one CPC code in G06F16 (information retrieval), G06F40 (natural language processing) or G06N (AI and machine learning) and has an English abstract. That is 2,170,848 publications in 1,214,491 patent families. For each publication we kept the English title, the English abstract and up to the first 20,000 characters of the English claims. English claims exist for 488,406 publications, mostly US, European and international (WO) filings, so text matching is deeper for those offices than for Chinese, Korean or Japanese filings, which we matched on title and abstract.

Within the 2,269 families that describe both generative AI and search, we kept the 358 whose text mentions ranking, relevance scores, or selecting sources, documents, search results, passages or web pages. We then flagged six signal groups by keyword: structured data (structured data, knowledge graph, knowledge base, schema.org, markup, metadata), popularity (popularity, clicks, click-through, engagement), location (geographic, geolocation, user location, proximity, nearby, GPS), freshness (freshness, recency, recently published or updated, publication date, up-to-date, time-sensitive), authority (authoritative, credibility, trustworthiness, source quality or reputation) and reviews (reviews, customer or user reviews, star ratings).

Structured data is the signal patents name most

We counted how many of the 358 families mention each signal group. A family can mention several.

Knowledge bases dominate this group because many filings describe answering from a company's own document store. Across all 2,269 AI search families, the structured data group appears in 824 (36.3%), while freshness appears in 46 (2.0%).

Freshness and authority are rising from a small base

We split the 358 families by first publication year.

Freshness went from 2 of 68 families (2.9%) in 2024 and earlier to 9 of 154 (5.8%) in 2026. Authority went from 1 to 7. The counts are small, so read these as early movement rather than a trend.

Most ranking patents do not spell out a signal

Many filings describe ranking with learned scores, such as similarity between a query and a passage, without naming business-style signals.

That fits how retrieval-augmented systems usually work: they score text by meaning first. Clear, relevant text on the page is likely the starting point for being selected.

What this means

The filings point to text relevance first, and to structured information second. For a business, that means pages should state what you do in plain words and back it with structured details: services, prices, hours, locations and FAQs marked up where it makes sense. Of the two, plain wording is the place to start, so AI search optimization services should fix the copy before the markup.

Freshness and reviews appear rarely in this patent text, but that does not make them unimportant for visibility in practice. Keep facts current and consistent. Some of your pages are already picked up by AI assistants; a free AI visibility audit shows which, and our AI SEO services add the markup and plain-text fixes the rest need.

Limits

A patent shows what was filed, not what is live in any product.

These counts describe patent text only. They do not show that any company ranks or selects sources on these signals.

Patents usually publish about 18 months after the first filing, so families first filed in 2025 and 2026 are heavily undercounted, and 2026 publication counts cover January 1 to September 17, 2026 only.

The structured data group includes metadata and knowledge bases, which often refer to internal document stores rather than website markup.

Signal matching is keyword based and deeper for filings with English claims.

Frequently asked questions

Related studies

• Patents behind AI searchQuery Fan-Out Patents: 44.1% Now Use Generative AI
• Patents behind AI searchSearch Patents 2026: 39.9% Now Mention Generative AI
• Patents behind AI search12.9% of AI Search Patents Use Personal History
• Patents behind AI searchChat Shopping Patents: 43.3% Now Use Generative AI