When a system tells you exactly how it scores you, the strategy writes itself. Google’s AI contribution pilot describes its scoring in three words: relevance, freshness, uniqueness. The research on which pages AI answers cite, gathered independently over two years, lands on the same three. So does the design of content generated from a knowledge graph.

This post takes each property in turn: what it means, what the evidence says, and what it implies for how a page should be built.

Relevance

What it means. How closely the content aligns with the specific intent of a query. Not the topic. The question.

The evidence. In a 252,000-trial controlled study across six language models this May, topical relevance and position in the retrieved list dominated every other factor, and formatting-only edits made no measurable difference (arXiv, May 2026). The 54-study synthesis published the same month ranked query-answer match, the answer appearing near the top of the page, and self-contained passages among the best-evidenced citation factors (Zyppy, May 2026). Google’s own guidance says no special markup is needed for AI features, and that the priority is “non-commodity” content (Google Search Central, updated July 2026).

Why it favors the long tail. People ask assistants specific, conversational questions: not “nursing schools” but “which nursing programs in Texas under twenty thousand a year have the best pass rates.” An assistant answering that question retrieves candidate passages that match it. A site with one broad page per topic has nothing that matches. A site with a page for every program, every place, and every cohort has a passage that matches almost any variant of the question. Relevance at the level of the query is a coverage problem, and coverage is what templates are for.

Design implication. Put the direct answer where a machine expects it: a self-contained passage near the top of the page that states the fact, the unit, the date, and the source. Structure the rest so each section stands alone. A passage that needs the paragraph above it to make sense is a passage that will not be lifted.

Freshness

What it means. Whether the facts are current.

The evidence. Pages cited in AI Overviews were about 26% fresher than the organic top ten across 17 million citations (Ahrefs, July 2025). Brands whose pages were not updated at least quarterly were three times as likely to lose citations, and only 30% of brands stayed visible from one answer to the next (AirOps, December 2025). Bing’s engineers describe their grounding index, the one that feeds Copilot and ChatGPT search, as scoring chunk-level evidence, provenance, and freshness (PPC Land, May 2026). Google’s June statement about its pilot named “freshness and factuality” as the qualities it is partnering for.

Why it is expensive for everyone else. A writer can add a statistic to an article. Keeping that statistic current across five thousand articles when the source dataset updates each year is a different problem, and it is where most content operations quietly give up. Content written by a language model has the same problem in a worse form: the model’s knowledge is frozen at training time, and refreshing means paying again while hoping the model has the new data rather than the old.

Why it is free for a data-native publisher. When pages render from a knowledge graph, freshness is a build, not a writing task. The data updates, the pages regenerate, the modified date is honest because something actually changed. A single data release becomes a freshness event across millions of pages in the same week, while competitors wait for writers.

Design implication. Show the data vintage on the page (“2024 cohort, IPEDS, released March 2026”), keep modified dates true, and regenerate on the schedule the data sets rather than the calendar. Make each release a small news event with a public changelog of what changed.

Uniqueness

What it means. Content that is not retrieved from other domains. In the pilot’s own words, a page scores lower “when multiple domains contribute similar information.”

The evidence. The original Princeton study on generative-engine optimization found that adding statistics raised visibility in AI answers by about 33%, quotations by 41%, and cited sources by 28% (arXiv, 2024). Google’s 2026 core updates rewarded originators and punished aggregators and reference databases under what Sistrix called “authority beats interchangeability” (Sistrix, March 2026). And earned media, the coverage that exists because someone had something new to say, accounts for 84% of AI citations across 25 million measured (Muck Rack, May 2026).

Why it is the property language models cannot fake. A model is very good at producing plausible general prose. What it cannot do is produce a verifiable, specific fact about the world that did not exist before. When an answer needs one, the model has to get it from somewhere, and that somewhere is what gets cited and paid. A page that restates the consensus adds nothing the model did not already have.

Why derived facts are the cleanest form of it. A ranking with a published method, a percentile within a cohort, a trend across years, a value score: these are computed from primary data and exist only where they were computed. They are unique by construction, not by effort. They are also honest, because they can be traced to the data and the method, which is the difference between uniqueness and novelty for its own sake.

Design implication. Every entity page should carry at least one fact that exists nowhere else, stated plainly with its method. Rankings and cohort comparisons are the obvious ones. “Where this school sits among its peers on graduation rate” is a unique fact. “This school has a good graduation rate” is not.

The same list, three times

	Google’s contribution scoring	The citation research	Content from a knowledge graph
Relevance	Alignment with the query’s specific intent	Query-answer match, answer near the top, self-contained passages	A page for every entity, cohort, and place
Freshness	Up-to-date facts	Cited pages 26% fresher; quarterly refresh or lose citations	Regeneration on every data release, honest dates
Uniqueness	Not retrieved from other domains	Statistics +33%, quotations +41%, sources +28%	Rankings, percentiles, trends, value scores computed from primary data

The table is the whole argument. What the platforms pay for, what the engines cite, and what data-native content has by construction are one list. That is why a shift that is bad news for most publishers is good news for the ones who own their facts.

What it is not

It is not schema. A controlled test that added JSON-LD to 1,885 pages found no uplift in AI citations (Ahrefs, May 2026). Emit clean structured data because it keeps the machine-readable layer honest, not because it earns citations.

It is not llms.txt. Google does not read it, and 97% of the files published across 137,000 domains received zero requests in a month (Ahrefs, June 2026). Publish one because it costs nothing; never bill for it.

It is not a byline on every page. Fabricated author profiles are rated Lowest in Google’s rater guidelines (Search Engine Journal, May 2025). Provenance comes from the method and the data vintage, and from a real, accountable person or an institutional desk that stands behind them.

Relevance, freshness, uniqueness. Build for those three, measure whether it worked, and let the theater go.