Content embedding alignment used to be something you talked about in theory and never actually measured. You could guess whether your page matched the meaning of a query inside an AI engine's head, but you had no way to check the guess. That changed this month. Dan Petrovic, managing director of DEJAN, published a free, 300-million-parameter model that predicts Google's Gemini embeddings at 0.83 cosine similarity, running locally, with no API key to Gemini itself required. That is a real, testable way to see how close your content sits to what Gemini considers the topic.
What Embeddings Actually Are (And What Keyword Density Never Told You)
Every AI model that ranks or retrieves content converts text into a list of numbers before it does anything else with it. That list is an embedding. It is the model's internal fingerprint for what a sentence, paragraph, or page means, built so that pieces of text with similar meaning land as numbers that sit close together in that space, and pieces of text with different meaning land far apart. When Gemini decides whether your page is relevant to someone's question, it is not scanning for your target phrase. It is comparing the embedding of the query to the embedding of your page and asking how close they are.
The distance between two embeddings is usually measured with cosine similarity, a number from 0 to 1 that tells you how closely two vectors point in the same direction. A score of 1.0 means identical meaning as the model understands it. A score near 0 means the model sees no relationship at all. This is the actual mechanism behind semantic relevance in AI search, and it is a different mechanism than the keyword matching that classic SEO trained everyone to watch. You can write a page dense with the right words and still sit far from the query in embedding space, and you can write a page with almost none of the exact phrasing and land close, if the meaning lines up.
The Content Embedding Alignment Problem Nobody Could Test
Here is the problem that has sat under generative engine optimization work since it started: embeddings are the mechanism, but the actual embeddings a proprietary model produces are locked inside that model. Gemini's embedding space is not published. You cannot inspect it directly unless you pay for API access and run your content and your competitors' content through Google's own endpoints at scale, which most marketing teams, and most agencies, never do. Everyone talks about optimizing for semantic relevance. Almost nobody has had a way to check their work against the model that actually decides it.
That gap produced a lot of vague advice. Write comprehensively. Cover the topic in depth. Use related terms. All of that is reasonable direction, and none of it is a measurement. A team could follow every piece of that advice and still have no idea whether their page landed anywhere near where Gemini places the ideal answer to a buyer's question. The alternative, before this, was building your own embedding model from scratch or paying for enough API calls to approximate Gemini's space yourself, and both of those put real embedding-level testing out of reach for the accounts that would benefit from it most.
This is also why so much of the debate around AI content optimization stayed stuck at the level of opinion. One writer swears by longer, more exhaustive pages. Another swears by short, direct answers. Both can point to examples that seem to prove their case, because without a way to measure alignment directly, every argument ends up resting on anecdote. A free, reasonably accurate proxy for Gemini's own embedding space changes the shape of that argument. Instead of debating which style sounds more thorough, you can check which version actually sits closer to the target meaning, for a specific query, on a specific page, today.
Inside DEJAN's Gemini Embedding Prediction
Dan Petrovic, managing director of DEJAN, published research around the middle of August showing a 300-million-parameter model that predicts Gemini's embeddings at 0.83 cosine similarity. Read that plainly: a small, open, free model anyone can download and run locally gets within 0.83 of a direct match to how Google's frontier model represents the same text. It does not require an API key to Gemini itself. It does not require paying Google per request. It runs on hardware a normal team already owns.
0.83 is not identical, and it should not be treated as identical. But it is close enough to be useful as a working proxy, which is the entire point. Before this, the honest answer to how does Gemini see this page's meaning was we don't know, and finding out would be expensive. Now the honest answer is we can approximate it, today, for free, and get back a number that means something.
That crawler figure is a side finding from the same research window, offered here only for context on how often these systems even come knocking. It is not the headline. The embedding prediction is the part worth changing your process over, because it is the first time the actual mechanism behind AI relevance, not a proxy for it, has been something a normal team can test directly.
Building a Semantic SEO Audit With a Free Model
A semantic SEO audit used to mean reading the page yourself and applying judgment, or leaning on tools that infer relevance from keyword co-occurrence, which is a proxy for a proxy. With a local embedding model that tracks Gemini at 0.83, the audit gets closer to arithmetic. Generate an embedding for your page. Generate an embedding for the query, for the ideal answer you're trying to match, or for a competitor's page that is winning citations you are not. Compare the vectors. The number you get back is a real signal about semantic distance, not a guess dressed up as one.
The output tells you something keyword tools cannot: not just whether the right words are present, but whether the meaning holds together at the level a model actually reads it. This pairs directly with how content gets broken into pieces and retrieved, because a page can score well as a whole document and still fail at the paragraph level, which is the level most retrieval systems actually operate on. Running the alignment check chunk by chunk, not only page by page, is where this stops being directionally interesting and starts being genuinely useful for AI content optimization work.
| OLD APPROACH | WHAT IT ACTUALLY MEASURED | EMBEDDING ALIGNMENT AUDIT |
|---|---|---|
| Keyword density / TF-IDF tools | Whether the right words appear, not what they mean | Whether the model's own representation of your text matches the target |
| Manual "does this sound thorough" review | A human's judgment of comprehensiveness | A numeric distance between your page and the target, in the same space Gemini uses |
| Paid API-based embedding comparison | A real Gemini embedding, at real cost, per call | A free local approximation at 0.83 similarity, no per-call cost |
| Related-terms / entity checklists | Presence of associated vocabulary | Semantic closeness regardless of the exact vocabulary used |
What 0.83 Similarity Does and Doesn't Prove
Two limits matter here, and both are worth saying plainly instead of glossing over. First, 0.83 is a strong correlation, not a perfect one. There is real daylight between the free model's prediction and Gemini's actual internal representation, and a page that scores well on the proxy could still score differently against the real thing. Treat every result as directional, not final. Second, this model approximates Gemini's embedding space specifically. It says nothing certain about how ChatGPT, Claude, or Perplexity represent the same content, because each of those systems builds its own space, trained its own way. A page that aligns well with the Gemini proxy is not guaranteed to align with what OpenAI's or Anthropic's models do internally.
Neither limit erases the value here. Directional signal, run consistently across your whole content set, beats no signal at all, which is what most teams have been working with. The right posture is to use this as a screening layer: run it across pages at scale, flag the ones that score furthest from the target, and save human review and real API-based verification for the pages where the decision has real budget riding on it. That mirrors how failure patterns in AI citations tend to show up in practice: the gap is rarely a single glaring error, it is a page that sounds right to a human editor but sits meaningfully off-target in the space a model actually uses to judge relevance.
It also helps to remember what this proxy is not. It is not a ranking algorithm, and a high alignment score does not guarantee a citation any more than a low one guarantees you'll be ignored. Retrieval, authority signals, and machine access all still sit between a well-aligned page and an actual mention in an AI answer. What this gives you is a missing diagnostic, not a finished scorecard. Treat a strong alignment number as necessary groundwork rather than a promise, and keep the rest of your measurement stack, the parts that track whether citations actually show up, running alongside it rather than in place of it.
Running Your First Content Embedding Alignment Audit This Week
This is a capability worth using immediately, not a research curiosity to file away. Most content and technical SEO teams have never had a way to check their AI content optimization work against anything other than intuition and a keyword tool. Now there is a free, running-today way to get a number instead of a hunch.
Pair the embedding check with the structural work that determines whether a model can even reach the right paragraph in the first place, since alignment only matters if the content is retrievable and well-formed to begin with. Teams running a full technical audit, especially B2B SaaS accounts where the buyer's question is narrow and specific, are best positioned to use this well: the audit tells you where you're misaligned, and a standing measurement practice, the kind covered in tracking citation and mention market share over time, tells you whether the fix actually worked. Download the model, embed five pages today, and see the number for yourself instead of guessing at it for another quarter.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.