[ tool ]

Agent Legibility Analyzer

Everything runs in your browser. Nothing you paste is transmitted or stored.No network request carries your text, there is no analytics, and nothing is logged. You can confirm this in your browser's network tab.

Paste a product payload (rendered HTML, a JSON-LD Product block, or plain description text) to see how a language-model agent would read it: how many tokens it costs, which decision fields are present, how much marketing noise dilutes the signal, and a leaner version that keeps only the facts.

Key takeaways

This tool runs entirely in your browser. Nothing you paste is sent to a server, logged, or stored.
Token counts are exact only for the OpenAI o200k_base encoding (GPT-4o family). Claude and Gemini publish no browser tokenizer, so their counts would be an approximation.
The legibility score is a heuristic self-assessment, not a measured audit. It rewards decision-field coverage, low marketing noise, and conciseness.
The lean view only removes marketing filler and reorders by decision weight. It never adds or infers data you did not paste.

Analyze a payload

Paste product HTML, JSON-LD, or plain text

Results update as you type. Everything stays local.

Paste a payload above to see its token count, structure breakdown, and legibility score. The methodology and a worked example are below and need no JavaScript.

The interactive analyzer needs JavaScript to run in your browser (that is how it keeps your text local, with no server round-trip). With JavaScript off, read the methodology and the worked example below to check a payload by hand.

Get new research and model-update notes

Agent selection changes with every model update. Subscribe for new findings on making product data legible to shopping agents. No spam, unsubscribe anytime.

Methodology

The analyzer computes four things from the text you paste, all locally. Each is a transparent, rule-based calculation, not a model prediction.

Token counts

Before the tokenizer loads, the tool shows an Estimate using a common rule of thumb of roughly four characters per token for English text. On your first analysis it loads the OpenAI o200k_base tokenizer (the encoding used by the GPT-4o family, which powers current ChatGPT Shopping) and shows an Exact count. That tokenizer ships its lookup tables inside this page's bundle, so it runs without any network request.

Anthropic (Claude) and Google (Gemini) do not publish a browser-usable tokenizer. A token count for those models would be an Approximation based on an OpenAI encoding, so the tool labels it that way and never calls it exact. Use token counts as a prompt-budget signal: leaner payloads cost fewer tokens to include in an agent's context.

Structure breakdown

The tool scans for seven decision-relevant fields that a shopping agent needs to compare and buy: price, currency, availability, GTIN or SKU identifiers, dimensions or weight, shipping terms, and return terms. Detection is a case-insensitive keyword and pattern match over the raw text, so it works on HTML, JSON-LD, and prose. It reports which fields are present and how many times each matched. Each field carries a fixed decision weight (price is weighted highest, dimensions lowest) that drives the lean view's ordering.

Marketing-noise indicator

The tool counts occurrences of a published list of marketing and puffery terms (for example "revolutionary", "cutting-edge", "world-class", "premium", "game-changing"). The full list lives in the tool's source so you can see exactly what is counted. The noise ratio is marketing hits divided by total words. High noise dilutes the decision-relevant signal an agent has to parse.

Legibility score

The score runs from 0 to 100 and sums three components. It is a heuristic self-assessment of machine-readability, not a measured ranking signal from any agent. The code matches this description exactly:

Decision-field coverage, 50 points. The share of the seven decision fields present, times 50. All seven present scores the full 50.
Signal purity, 30 points. Full at zero marketing noise, falling to zero once marketing terms reach 5 percent of total words.
Conciseness, 20 points. Fewer words per detected field is leaner: about 40 words per field earns the full 20, tapering to zero near 400 words per field.

Token counts are shown for prompt-budget context and are deliberately not a score input, so the score does not depend on which model's tokenizer you compare against.

Lean view

The lean view is a de-noised, decision-ordered outline of the exact text you pasted. It does two things and only two: it removes segments that are pure marketing (a segment is dropped only when it is prose carrying a marketing term but no decision field and no number, so structured data is never removed), and it reorders the remaining segments by decision weight. It never rewrites, invents, or infers a value you did not paste. Treat it as a suggestion, and review it before you use it.

Limitations

Read this before trusting the number

The legibility score is a heuristic self-assessment, not a measured auditand not a ranking any agent publishes. A high score means the payload is clean and complete, not that a specific agent will select it.
Token counts are exact only for OpenAI o200k_base (and cl100k_base for legacy GPT-4 and 3.5). Claude and Gemini counts are approximations, because those vendors ship no browser tokenizer.
Field detection is keyword and pattern based. It can miss data expressed in an unusual way and can over-count a word used in a non-product sense. Read the breakdown, do not take it as a schema validator.
This tool checks legibility, not eligibility or correctness. It does not validate your structured data, crawlability, or protocol setup. For those, use theAgentic Shopping Readiness checklist.

Worked example

Illustrative example, not measured data.

A noisy product blob buries the facts an agent needs under marketing language. The lean version keeps every real value and drops the filler.

Before: noisy

Introducing the revolutionary, world-class AuroraGlow desk lamp, a truly
game-changing addition to any premium workspace. This stunning, best-in-class
lamp is unlike anything else on the market. Elevate your desk with its sleek,
elegant design. Price: $49.00 USD. In stock. SKU: AUR-DL-200. Ships in 2-3
business days. Free returns within 30 days. Dimensions: 42 cm tall, 1.2 kg.

After: lean

Price: $49.00 USD. In stock. SKU: AUR-DL-200. Ships in 2-3
business days. Free returns within 30 days. Dimensions: 42 cm tall, 1.2 kg.
AuroraGlow desk lamp.

The lean version drops three sentences of pure marketing and keeps price, currency, availability, identifier, shipping terms, return terms, and dimensions. Same facts, far fewer tokens, and the decision-relevant fields lead. Paste your own payload above to run the same pass on it.