Botify Study: Does Semantic Similarity Predict AI Citation Likelihood?

What did the Botify study find?

Botify ran the study with DemandSphere, a search analytics company, and released it in December 2024. Its public findings:

The full report, including the similarity scores and method, is behind a form, and public summaries give no numbers for the similarity finding. DemandSphere later built an AI Overview cosine similarity feature into its keyword research tool on the back of the analysis.

What does semantic similarity measure?

Semantic similarity compares meaning rather than exact words. Each text is converted by an embedding model into a list of numbers, and cosine similarity measures how closely two of those lists point in the same direction. Texts about the same thing score high even if they use different words.

Two points matter for reading any study that uses it:

• Scores depend on the model. Different embedding models produce different ranges. A score of 0.85 from one model is not comparable with 0.85 from another, so there is no universal threshold.
• Scores cluster. Texts on the same topic often all score high. In a separate analysis of product descriptions in AI answers, G2 found similarity of about 0.83 to the product's own profile, only around six points above a baseline comparing different products. Small differences carry the signal, and they are easy to over-read.

Does the finding show that similarity predicts citation?

It is weaker evidence than it sounds, for three reasons:

1. It measures similarity to the answer, not the query. An AI Overview is generated from its sources, so the sources will resemble it. That tells you which pages contributed, not why they were chosen.
2. It does not separate similarity from ranking. Cited pages also ranked highly, and highly ranked pages are already the most relevant to the query. Public summaries do not show whether similarity adds anything once ranking is accounted for.
3. It predates big changes. Since January 2026, Gemini 3 has powered AI Overviews, and Ahrefs' 2026 update found only about 38% of AI Overview citations came from the top 10 results, down from 76% in its 2025 study. The late-2024 top-12 figure is probably too high for today.

The practical reading still holds up: pages that directly and fully answer what the Overview says are the ones it links to. But “similarity predicts citation” is an inference, not the study's measured result.

What does later research add?

An April 2026 paper on arXiv studied 602 prompts and more than 21,000 citations across AI search platforms, and split the problem in two:

• Citation selection: whether a page is chosen as a source. This depended on authority, recognisability, language and domain context.
• Citation absorption: how much a chosen page shapes the answer. High-influence pages were longer, more modular, more semantically aligned with the answer, and richer in extractable evidence such as definitions, figures, comparisons and steps.

The two outcomes diverged by page type. News pages were selected often but, on the paper's influence measure, averaged about a third of the influence of encyclopedia pages once cited.

The authors warn that measures such as text overlap and similarity to the answer are partly how influence is defined, so they cannot be treated as independent causes of it. That is the same circularity that affects the Botify finding. Word count and similar signals are discussed in whether word count affects AI citations.

Where does the “0.88 cosine means 7.3 times more citations” claim come from?

The figure appears across many GEO blogs as a hard rule. Following the references back, it is attributed to a vendor blog that cites an analysis of 15,847 AI Overview results but gives no method, no embedding model and no link between that dataset and the 0.88 figure. Since cosine scores depend on the model used, a single threshold such as 0.88 cannot apply across tools. Treat it as unverified, and be wary of tools that promise a target similarity score.

What should you do with the findings?

1. Get into the candidate pool first. Indexing, crawl access and ranking still decide which pages are considered. Similarity cannot help a page that is never retrieved.
2. Answer the question in plain terms early, using the words people use, then cover the follow-up questions a search might branch into. Do this within one strong page, not a separate page for each variation, which Google treats as scaled content abuse when the pages are made mainly to manipulate rankings, as covered in whether scaled content can be penalised.
3. Package evidence so it can be lifted: definitions, numbers with sources, comparisons and steps.
4. Use similarity tools for diagnosis only. A low score against an Overview can reveal a page that answers a different question. Do not stuff related terms to raise a number.
5. Measure citations directly for your key questions, rather than trusting a proxy. Record which page is cited, not only whether your domain appears, so you can see which pages the answers actually use.

Botify's broader platform is assessed in the Botify agentic AI platform review.

We have not run a similarity study of our own. This page is based on Botify's and DemandSphere's published summaries, Ahrefs' 2026 update, G2's analysis and the April 2026 arXiv paper.

What might this look like in practice?

Consider an illustrative garden supplies retailer, Thornbury Growers, that bought a tool scoring pages by cosine similarity to AI Overviews and set a target of 0.88 for every page. The team began adding lists of related terms to category pages, and scores rose while citations did not. A review found the pages were not ranking for the questions at all. They stopped chasing the score, rewrote the five most important guides to answer their questions in the opening paragraph with specific figures and steps, and fixed internal links. Two of the five began appearing in AI Overviews over the following two months.

What do people ask about semantic similarity and AI citations?

Does higher cosine similarity guarantee a citation?

No. It is one weak signal, and pages must first be retrieved and trusted.

What cosine similarity score should I aim for?

There is no universal target, because scores depend on the embedding model used.

Did Botify prove semantic similarity causes citations?

No. It found cited pages resemble the AI Overview's text, which is partly expected because the Overview is written from them.

Is ranking still important for AI Overviews?

Yes, but less than it was. Citations now come from a wider range of positions than in 2024.