What the AI citation studies agree and disagree on

Cyrus Shepard shortlisted 55 studies, experiments, explainers and patents on AI citations, and then used them to score 23 possible factors¹. I read the methods of each study and the numbers within the paper rather than the summary score. I am trying to show the sample sizes on both sides of each disagreement, so you can judge them against your own situation.

Keep in mind that correlation is baked into every study here. Most of the study authors are selling AI visibility tools. None of the 55 are legal studies, but surely at least some of the findings apply to law firms? That doesn't make them useless, but I hope it helps to read them as evidence about the web in general. Web-wide evidence has to hold in a law firm's own market to be useful there.

This page is just the evidence appendix for the rest of the section, and it is really practical too. When pitched a package that has a study attached you can look up the method and sample behind it. Finding a study of similar size that says the opposite is more common than you think. It happens on two of the questions below (the ones firms ask first).

Six findings the studies agree on

1. Ranking in ordinary search is the strongest single factor

In seoClarity's analysis (362,000 US queries that produced an AI Overview), 94% of the answers were sourced (with at least one cited page) within the top 20 organic results².

ChatGPT cited 43.2% of the pages ranking first in Google, and AirOps has the data on what was and wasn't cited. Pages outside the top 20 have roughly 3.5 times fewer citations than first-place pages³.

2. The answer has to sit near the top of the page

One study took 42,971 citations that it traced back to their source pages on six platforms. The median depth of a cited sentence is 31.2% of the way down the page. The 25th-percentile sentence of the set sits at 18%. So if three quarters of the sentences that get cited sit in the first half of the page, we say that's where the answer goes⁴. As a grounding study, Dejan's work sees this from the other side. In those tests the model has almost completely lifted the opening paragraphs, skipping the unrelated sections⁵.

3. Headings that match the question get cited more

It turns out you should word your headings closely to the query, which lined up with a citation 41% of the time across 16,851 queries and 353,799 pages⁶. Pages with 50% or more overlap between title and query command a 20.1% citation rate. That's 2.2 times as many citations as pages under 10% overlap actually got (9.3%)³.

4. Specific facts beat general statements

In its test, Surfer matched 57,000 URLs against the AI Overviews for 1,591 keywords and worked out what share of each topic's key facts every page covered. The average fact coverage of a cited page is 31%, versus 24% for an uncited page. The median cited article is 62% more complete than the typical uncited one⁷. Pages with five to seven supporting statistics are getting cited more (a 20.3% higher likelihood across 217,508 pages and 7,500 commercial prompts)⁸.

5. Structure helps extraction

The 42,971-citation study found that of its 1,931 usable source pages, 95.6% had at least one list, table or set of multi-level headings⁴. Comparison pages with three tables earn 25.7% more citations. Shortlist pages with short sentences (ten words or fewer on average) earn 18.8% more⁸. SE Ranking has found that across 129,000 domains, pages with sections of 120 to 180 words between headings draw about 70% more ChatGPT citations, with the comparison group at under 50 words a section⁹.

6. Schema markup has a small but consistent effect

Structured data is rated 5.6/10 across the collection and has a small but consistent effect from study to study¹. After controlling for domain authority, schema-marked pages got 2.3 times as many citations as unstructured ones in the analysis of 1,000 AI Overviews¹⁰.

The studies disagree about how long a page should be

This is the biggest contradiction of the entire set. Samples on both sides are more than large enough to take seriously.

Provided the page is long enough to cover the question's follow-ups and you put the answer at the top you don't even need any extra length to compete. Usually 1,000 to 2,000 words on a practice area page. Which is why I try to avoid writing one enormous guide and rather split it into twenty pages instead.

The studies disagree about backlinks and domain authority

It's the same thing with links as with length, and, I imagine, with domain authority as well. It's links that earn the rankings that put a page in the retrieval set, and from there it's the text that decides. Studies across the whole web, with weak and poorly ranked pages included, get a fair bit of link effect. Chiefly because links got those pages to the starting line. None of the studies of already-retrieved pages pick it up. When a study starts at the retrieval set the links had already done their work before the study began.

The second factor is what Dejan calls primary bias. Nothing to do with links at all. Primary bias is the term he uses for the model's existing sense of how relevant a brand is to a subject. That sense is formed in training and fine-tuning and Dejan says you'll see it shift over three to six months rather than within a campaign¹². Brand and entity trust is higher scoring across the collection (6.8) than domain authority¹.

The studies disagree about freshness

Freshness matters a lot (7.0 of 10) for questions about recent events¹. But the median age of a page cited across 1,000 AI Overviews is 14 months¹⁰. The main problem I see many firms making is relying solely on a new date to look fresh while having no other change to the page. A new date should almost always be a part of a real update to the page after the law or the market changes.

How concentrated AI citations are, and how much that varies by platform

Roughly twelve domains receive about 47% of the citations in AI Overviews, 22% goes to the long tail, which is less than half of what those twelve take. The twelve include Wikipedia, Reddit, Forbes, Healthline and the larger government and education sites. Only 8% of AI Overviews in that data cite more than seven domains. For reference, the median number of cited domains is four for each Overview¹⁰. ChatGPT's citations include just 12% from its top ten domains and the rest from a far wider spread. Really, it's just the inequalities of a broad market (a Gini coefficient of 0.8, where 1 would mean total concentration)¹³.

The other thing is platform, and how closely each one follows Google. In both the domain and the URL counts, Perplexity is the most Google-like platform, matching 91% and 82% of Google's top 10 respectively. Google's own AI Overviews overlap around 86% by domain and roughly 67% by URL. AI Mode is Google's too and has about 54% domain and 35% URL overlap¹⁴. Only 19% of the citations out of seoClarity's AI Mode sample come from the top 20 organic results¹⁵.

Against a page-quality framework, the platforms differ in the quality of pages cited and the rate they cite them at. Brave first at 78%, then Google AI Overviews at 72%, then Perplexity at 45% with markedly lower quality¹⁶.

What this evidence means for a law firm specifically

I'm aware that in legal, a Your Money or Your Life subject, AI answers rely heavily on ordinary rankings. BrightEdge took a deep dive into AI Overview citations over sixteen months, and the overlap with organic results climbed from 32% to 54%. Trust-heavy categories (the YMYL ones) are at 68 to 75% overlap. BrightEdge says Google strongly connects its trust-heavy answers to content that already ranks well¹⁷.

So most of the AI visibility work for a firm is done on ordinary rankings for the questions clients ask. You need to answer the client's question within the first paragraph, if not sentence. But 80.9% of website citations that came up in one analysis have been pointing at third parties instead of the brand¹⁸. Getting cited on third-party sites has its own page: off-site work.

For smaller firms, there is one finding here that matters most. Reciprocal rank fusion is a scoring method for how ChatGPT appears to combine results from related searches. Since any ranking for any related query flows straight into the fused score, ranking around sixth for thirty of them would imply a roughly 10x better score than first for three¹⁹. I spread a firm's effort across the whole practice area.

What Google itself says, which is less than agencies claim

"There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." And: "You don't need to create new machine readable files, AI text files, or markup to appear in these features."²⁰

The only thing that Google documents as a requirement is a page that is indexed and eligible to show with a snippet. A nosnippet page or one that blocks the crawler is excluded and dropped quietly, so these AI answers never see it²⁰. Shepard cannot see any evidence to substantiate the llms.txt claims. Rated 2.0/10, over 55 studies reviewed, still sold to firms as AI optimisation¹.

Four reasons to hold all of this loosely

• The evidence is correlational. Shepard's report does say this plainly. There is a difference between a plausible ranking factor and showing actual causation¹.
• Most of it is vendor research. Run by the people who sell visibility tools, these studies are, again, usually without a shared dataset.
• None of it is legal. Every figure on this page comes from general web data and from other industries.
• The platforms move faster than the research. ChatGPT's citation rate for Reddit fell from roughly 60% of responses to about 10% within weeks, while Wikipedia's fell from about 55% to under 20% there. Other platforms, on the other hand, kept both steady²¹. So the measurement research's first recommendation is to sample visibility across many runs before trusting any number²².

I'm working on a citation check of legal queries to help replace the borrowed numbers with ones from this market. But those results are not finished with yet.

How I use this when somebody quotes a study at me

I've seen plenty of pitches written by people who could answer the first of these four questions, but couldn't answer the second or third.

1. What did they actually measure? Those four metrics (citation rate, retrieval rate, visibility share and overlap with organic) are completely different even when they seem to be related. It's trivial to switch between a citation rate and a visibility share in a vendor deck. The grounding and citation studies (measuring different stages of one pipeline) are not mutually exclusive, they should go hand in hand if you're going to read either properly.
2. Did they control for rank? Most of the contradiction on this page is that you're comparing studies that never held rank steady. If you know rank is the strongest factor in the whole set, and a study doesn't hold it steady, then its result is more likely a rank effect. Wasn't this exactly what the link disagreement was, though?
3. Is there a study of similar size finding the opposite? Yes, I've found two actually (page length and backlinks, both set out above with their samples). A one-sided vendor study says things that are strictly true and presents its findings in the best possible light, while leaving off anything that might balance out the comparison.
4. Who paid for it, and what do they sell? That doesn't make it disqualifying, but it's a very big clue to which way the thumb was on the scale. Vendor studies of the thing their own tool improves deserve and need every bit of scrutiny on method that they can get.

Most of what I act on in this area is the narrow set that survives every one of those questions. Rank for the question, answer it near the top of the page, make the headings match, be specific enough to quote, keep the page open to crawlers. Most of that list seems pretty common sense, as long as you're already doing good SEO. I honestly just think that's the fairest summary of the field at the moment.

I avoid llms.txt like the plague. Same for AI-specific files or markup schemes, packages that promise AI Mode results, brand thresholds quoted at a local firm from studies of national publishers. You can see the reasoning for each on the AI search page and the authority page.