Real Retrieval Simulator — Free by AEO Insider Skip to content

Free Tool — Runs in Your Browser, No AI Account Required

Which Paragraphs Would AI Actually Quote?

Word-count rules and banned-phrase checklists can’t tell you that. This tool can — it runs a real embedding model, the same kind of technology behind retrieval-augmented generation, directly in your browser, and shows you which of your paragraphs would genuinely surface for realistic reader questions.

Real Retrieval Simulator

See which of your paragraphs an AI system would actually retrieve. This runs a real embedding model in your browser and measures genuine similarity between your content and realistic reader questions — not a word-count guess.

Your content

Works best with 3–15 paragraphs. Very short input won’t produce a meaningful simulation.

Candidate questions

Generated from the terms that repeat most in your text. Edit freely, delete ones that don’t fit, or add your own — one question per line. Better questions here make the whole simulation more useful.

Working…

Retrieval Coverage Score

0 of 0 questions have a strong matching paragraph.

Retrieval heatmap

Weak matchStrong match

Candidate questions are generated with a simple keyword heuristic, not an AI model — edit them before running if you want a sharper simulation. Weak-chunk flags are pattern-based (dangling references, word count) rather than full linguistic coreference resolution. Both are documented in the README, along with how to swap in an LLM for higher-quality questions later if you want to.

Methodology

How the Retrieval Coverage Score Is Calculated

This isn’t a readability formula or a word-count check. It’s a real, reproducible measurement of vector similarity.

Chunking

Your content is split into paragraph-sized pieces, targeting the 50–120 word range that retrieval systems generally handle best.

Embedding

Every chunk and every question is converted into a vector using all-MiniLM-L6-v2, an open-source embedding model. This runs entirely in your browser.

Similarity Scoring

For every question, we measure its cosine similarity against every paragraph — the same underlying mathematics a RAG system uses to decide what to pull into an answer.

Coverage Scoring

A question counts as “covered” if at least one paragraph scores above the strong-match threshold. Your score is the share of questions that clear that bar.

AEO Insider Editorial Team Updated September 2026 · 11 min read · GEO & AEO

Why Embedding-Based Retrieval Is Different From Word-Count Chunking Advice

The common advice for AI-friendly writing is a set of rules of thumb: keep paragraphs between 50 and 120 words, avoid phrases like “as mentioned above,” put one idea per paragraph. These rules exist for a real reason — but following all of them doesn’t guarantee a paragraph will actually be retrieved for a given question, because retrieval isn’t governed by word count. It’s governed by meaning, measured as distance in a vector space.

How to Use This Tool

Paste your content. Separate paragraphs with a blank line — the tool uses those breaks to find natural chunk boundaries, the same way a retrieval system would.
Click “Suggest candidate questions.” The tool scans your text for its most repeated terms and turns them into plausible reader questions. This step is instant and free — it doesn’t call any AI model.
Edit the question list. Delete anything off-topic, and add real questions your audience actually asks, in their own words. The simulation is only as good as this list.
Click “Simulate retrieval.” The first run downloads a small embedding model to your browser (a few seconds, one time only); every run after that is fast.
Read the heatmap. Rows are your paragraphs, columns are your questions. Darker cells mean a stronger, more retrievable match.
Fix what the Coverage Gaps and Weak Chunks panels flag. Gaps show questions nothing on the page answers well. Weak chunks show paragraphs with a dangling reference or an awkward length.
Re-run after editing to confirm your changes actually moved a question from red to green — not just that you added more words.

A note on the first run

The first simulation on a new browser downloads a small (~25MB) open-source embedding model. Your browser caches it after that, so every simulation afterward — even on a different page — is fast.

A perfectly-sized, self-contained paragraph can still score poorly against a specific question if its wording drifts from how that question is actually phrased. And a paragraph that breaks a rule of thumb can still retrieve well if its meaning is close enough to the question. This tool measures the thing that actually determines retrieval — similarity — rather than the proxies that are easier to check by hand.

How to Interpret Your Retrieval Coverage Score

Score	What it means	Typical profile
80–100	Strong retrieval coverage	Most realistic reader questions have a clearly matching paragraph
55–79	Solid but uneven	Core questions are covered; some specific angles are thin
30–54	Patchy coverage	Content touches the topic broadly but rarely answers a specific question directly
0–29	Weak coverage	Content and likely reader questions are largely disconnected in meaning

A low score usually means one of two things: the content genuinely doesn’t address certain angles yet, or it addresses them using very different language than a reader would use to ask about them. The heatmap tells you which.

The Priority Order for Fixing Retrieval Gaps

Close the reddest coverage gaps first A question with no matching paragraph at all represents a completely missed angle — that’s a bigger loss than a paragraph that’s merely a little weak.

Match the reader’s phrasing, not just the topic If a paragraph covers the right idea but scores poorly, try rewriting it using the same words and framing the question itself uses, rather than more formal or internal terminology.

Fix dangling references before rewriting anything else A paragraph flagged for opening with “This approach…” or “As mentioned above…” will underperform regardless of how good its content is, because it isn’t self-contained. Name the subject directly.

Split paragraphs that are carrying two ideas A long paragraph often scores moderately against several questions instead of strongly against one — breaking it up usually raises both scores.

Re-run before moving on Confirm a fix actually raised the specific cell you were targeting in the heatmap, not just the overall score.

Why This Runs in Your Browser Instead of Calling an AI Model

Most retrieval-adjacent tools either skip real embeddings entirely (falling back to word-count rules) or send your content to a paid AI API to approximate this kind of analysis. This tool takes a third path: it runs an actual, open-source embedding model client-side, using WebAssembly.

No account, no key, no per-use cost. The model is free and public; your browser downloads and caches it once.
Your content never leaves your browser for the embedding step itself — only the optional “fetch from URL” mode touches your own server, and nothing is sent to a third-party AI provider at any point.
Results are reproducible. The same content and questions will produce the same similarity scores every time, because it’s a mathematical measurement, not a generated response.

What This Tool Doesn’t Measure

It doesn’t generate publication-ready questions. The candidate questions come from a keyword heuristic, not an AI model — they’re a useful starting point, and editing them before running is expected.
It doesn’t perform true linguistic coreference resolution. Dangling-reference detection is pattern-based; it will catch common phrasings and can miss subtler ones or occasionally flag a paragraph that’s actually fine.
It isn’t a prediction of any specific AI engine’s behavior. The embedding model used here is a small, general-purpose one — a credible proxy for retrieval behavior in general, not an exact simulation of ChatGPT’s, Perplexity’s, or Google’s internal retrieval systems.
It doesn’t assess writing quality, factual accuracy, or SEO fundamentals like backlinks or domain authority. This is a single, specific signal: would this paragraph surface for this question.

Using This Tool Alongside the Rest of Your Content Workflow

Retrieval and topical coverage are two different problems, and this tool only solves one of them:

Confirm topical coverage first — make sure your content addresses the concepts a knowledgeable source on the topic would be expected to cover, using the Entity & Knowledge Graph Gap Mapper.
Then run the Real Retrieval Simulator to check whether the paragraphs you’ve written would actually surface for realistic reader questions.
Fix coverage gaps and weak chunks, prioritizing the reddest cells in the heatmap.
Re-run both checks after major edits — a rewrite that improves entity coverage can still leave retrieval gaps, and vice versa.

Frequently Asked Questions

What is retrieval simulation?

Retrieval simulation measures how closely your content’s meaning matches realistic questions a reader might ask, using the same vector-similarity approach that retrieval-augmented generation (RAG) systems use to decide what to pull into an AI-generated answer. It shows you which paragraphs would likely be retrieved, and which wouldn’t.

How accurate is the Retrieval Coverage Score?

It’s a real, reproducible similarity measurement using an open-source embedding model — not a heuristic guess. It’s accurate as a general proxy for retrieval behavior, but it isn’t a guaranteed prediction of what any one specific AI search engine will do internally, since each uses its own model and ranking logic.

Why does a paragraph I think is great score poorly?

Usually because its wording differs from how the matched question is phrased, even if the underlying idea is the same. Try rewriting the paragraph using language closer to the question, or check whether a dangling reference is quietly disconnecting it from its own subject.

What’s the single most impactful fix?

Closing complete coverage gaps — questions with no strong match anywhere — matters more than polishing a paragraph that’s already scoring moderately well. Start with the gaps list, not the heatmap’s weakest cells.

Does a high score guarantee my content will be cited by an AI system?

No. It means your content is well-positioned for the retrieval step specifically. Whether it’s ultimately cited also depends on factors this tool doesn’t measure: competing content, domain trust, and how closely your specific phrasing matches a real query.

Is my content sent to any AI model?

No. The embedding step runs entirely inside your own browser using an open-source model — nothing is sent to ChatGPT, Gemini, Claude, or any other AI provider. If you use an optional URL-fetch mode, that request goes to your own server, not a third party.

Why is the first run slower than later ones?