Embeddings

Do two questions that mean the same thing get the same answer?

Not necessarily. Two questions we track for attensira.com are near-synonyms: "AI visibility platform" and "AI search visibility software". An embedding model would place them almost on top of each other. The table at the top of this page shows what ChatGPT and Perplexity did with them over the 30 days to 4 October 2026, read over the Attensira MCP.

The headline held. Semrush was the most-named brand on both, in 50 of 106 answers and in 63 of 109. Underneath, the sources split. ChatGPT's most-cited source for "platform" was elmohq.com, a competitor's site, with 84 of 1,962 citations. For "software" it was technologyadvice.com, a software-review site, with 71 of 2,006. Neither domain was among the other question's six most-cited sources. Attensira was named in neither: 0 of 106 and 0 of 109.

We do not know how ChatGPT retrieves, so we cannot say which step split them. "Platform" may match vendor pages that call themselves platforms; "software" may pull the search towards review sites that list software. The lesson for anyone writing pages holds either way: meaning-based matching is close, not identical, and the exact word a buyer uses still decides which pages are read.

What is an embedding?

A list of numbers, typically hundreds or thousands long, that an embedding model produces for a piece of text. The model is trained so that texts with related meanings produce vectors pointing in similar directions. "Cheapest plan" and "entry-level price" end up near each other; "cheapest plan" and "crawler logs" do not.

Software compares vectors with simple arithmetic, usually cosine similarity, which makes it possible to search millions of passages by meaning in milliseconds.

Sizes vary by model. OpenAI's text-embedding-3-large returns 3,072 dimensions by default and text-embedding-3-small returns 1,536, and both can be shortened with a dimensions parameter.

How do AI answer engines use embeddings?

As the first sort in retrieval:

1. Pages are cut into passages; see chunking.
2. Each passage is embedded and stored in an index built for vector search.
3. When a question arrives, it is embedded with the same model.
4. The passages whose vectors are closest to the question's are retrieved, often merged with keyword matches from BM25, and passed to the language model to read.

This is the retrieval half of retrieval-augmented generation. It is why a page can be found for a question it never phrases exactly, and why semantic search beats pure keyword matching on paraphrase.

Where do embeddings fall short?

• Exact strings. A product code, an error number or a version can be missed while a page that is merely "about" the topic scores higher. That is why most retrieval systems pair embeddings with keyword matching; Anthropic's measured gain from doing so is on BM25.
• Context loss. A passage that says "it" instead of the product name embeds as something vaguer than the author meant.
• Model lock-in. Vectors from one model cannot be compared with another's. Change the model and the whole collection must be re-embedded.
• Near is not same. Our two near-synonym questions show it: the same leading brand, different sources underneath.

What does this mean for writing pages?

Write passages whose meaning is unmistakable on their own, and use the words buyers use. Put the product name and the exact terms inside the passage, so both meaning-based and keyword matching can find it. And cover the main phrasings of a question, such as "platform" and "software", in the page that owns it, instead of creating a second page for each.

What are the objections to writing for embeddings?

"Embeddings understand everything, so wording doesn't matter." Our table says otherwise.

"We should embed our site to see how AI sees it." You would see how your chosen model sees it. Each engine uses its own models and indexes. Reading their answers is the only direct evidence.

"Keyword research is obsolete." It changed shape. The phrasings buyers type into AI assistants are still the starting point.

How does Attensira use this?

Attensira tracks several phrasings of the questions that matter to you and reads each on every engine, so a split like our platform and software pair shows up as data rather than a guess; prompt tracking covers how phrasings are added and read. The crew then makes the page that owns the question answer every phrasing directly, ships the change after your approval, and measures each phrasing before and after. That is the service of an SEO agency at the price of a tool, applied one phrasing at a time.