All rules here are best practice, not law. Override any of them when the campaign calls for it; note the best practice once and move on.
Use when: the campaign needs to separate self-serve companies from sales-led ones, or needs a real price point to anchor an offer against.
Do not use when: you want technology installed on the site (playbook-tech-on-website).
One-line output: pricing_public = true with
plans = [{"name":"Basic","price":"$10","period":"month"}, ...] and lowest_paid_price = "$10".
1. Trigger and scope
Does this business publish real prices on its own website, and if so what are the tiers?
It is used for filtering first and copy second. The most common job is splitting a list into product-led companies (public price list, card checkout, low ACV) and sales-led companies (every tier says Contact Sales) — because those two groups need completely different offers and completely different copy.
Bones only by default. No pricing line unless the campaign asks.
The deliverable is the structured record: pricing_public, pricing_model, plans,
lowest_paid_price, has_free_tier, enterprise_quote_only, feature_diff_axis.
The one-clause string pricing_line is an opt-in extra, it is the weakest part of the output,
and it should be stripped at the output boundary unless explicitly requested. A price in a
stranger's inbox is a claim you have to be right about, and most campaigns using this playbook only
ever wanted the filter.
It does not cover: what a prospect actually pays after negotiation, private rate cards, marketplace listings, or ecommerce SKU prices.
If every tier says Contact Sales, the correct answer is pricing_public: false — a real and
useful signal, not a failure.
2. Output contract
Inputs required per row
| Field | Type | Required? |
|---|---|---|
domain (bare, lowercase, no scheme, no www) | string | yes |
company_name | string | no, improves the copy line only |
Output fields
| Field | Type | Example | Null? |
|---|---|---|---|
pricing_public | boolean | true | no, defaults false |
pricing_url | string | https://www.notion.com/pricing | yes, "" |
pricing_model | per_seat / usage / flat / credits / quote_only / unknown | per_seat | no |
plans | array of {name, price, period}, max 6 | see above | yes, [] |
lowest_paid_price | string | $10 | yes, "" |
has_free_tier | boolean | true | no |
enterprise_quote_only | boolean | true | no |
feature_diff_axis | ≤6 words | seats and automation limits | yes, "" |
confidence | high / low | high | no |
pricing_line | OPT-IN, off by default | your Basic plan is $10 per user per month | absent on a default run |
⚠️ quote_only and unknown are not the same thing
quote_only— the page exists and every tier says contact sales. A confident finding you can filter on.unknown— you could not read the page at all. A failed fetch, not a finding. The row must not be filtered out on your say-so.
Collapsing these two is how a campaign silently deletes real prospects because of a network error.
Coverage expectation
8/10 rows produced a usable filtering record. The existence check alone was 10/10. That is the number to plan on, because the filtering record is the whole default deliverable.
The optional copy string was usable on 9/10 in the clean regrade but 6/10 once corrected on an earlier run — so treat it as a 70-to-90% field: spintax-optional, never mandatory. That spread is part of why it is off by default.
Downstream gate
Whether a pricing value removes a row is decided per campaign brief. This playbook produces the bones; the brief says what to do with them. Ask once, up front:
- Filter (
pricing_public: falserows leave this campaign) — correct when the offer only makes sense for self-serve companies, or only for sales-led ones. - Route (send those rows to a different campaign or offer) — correct when both halves of the list are sellable, which is the common case.
⚠️ Either way, pricing_model: "unknown" never removes a row. Only act on confidence: "high"
findings.
If the brief is silent, route rather than delete, and say so in the handoff. Deleting rows is a client-facing, money-adjacent decision, and it is cheap to reverse a routing choice and expensive to reverse a deletion.
3. Source chain (cost-tagged)
| # | Source | Cost | Call | Result |
|---|---|---|---|---|
| 1 | URL-guess existence check | FREE | HEAD https://{domain}/pricing, then /pricing/, /plans, /pricing-plans, /price, then the same on www.. Browser UA, follow redirects, and accept only when the final status is 2xx AND the final resolved path still matches /pricing|plans|price/ | 10/10 correct. 15/16 pages found on the first candidate — but the candidate walk is not optional: one company only resolves via /pricing → /plans/, and two needed all 7 |
| 2 | Plain GET of the found URL | FREE | same browser UA, strip script/style/svg/comments to visible text, then take the price-dense window (not the first N chars), capped at 14,000 characters | 15/16 pages readable |
| 3 | Rendering proxy | METERED | JS render with a wait | 0/1 rescued. Only when step 2 produced under 800 characters of visible text OR fewer than 2 currency amounts. Never otherwise |
| 4 | SERP site search | CHEAP | site:{domain} pricing, take the first result on the company's own domain, loop back to step 2 | mixed. It located the real pricing URLs for two hard domains but returned docs/community/careers pages for two others. A discovery aid, not a guarantee. One simple query, no OR chains |
| 5 | Model extractor | CHEAP | reasoning_effort: "low", max_completion_tokens: 6000, JSON response format | 6/8 readable pages fully correct |
If HEAD is refused, fall back to GET with Range: bytes=0-4095 — and accept any 2xx, because
a compliant origin answers 206 Partial Content, not 200. A script that accepts only 200 silently
records those domains as having no pricing page.
When "pricing analysis" is bigger than this chain
This playbook answers a narrow, mechanical question at 8/10 for a fraction of a cent, so do not route the per-row extraction to an agent of any kind.
But operators say "pricing analysis" for a second, genuinely different job: work out how this company actually charges and what that implies for our pitch — packaging logic, where the value metric sits, how it compares to a competitor, whether "contact sales" hides a floor. That is a research question, and for it a strong reasoning subagent per account generally beats a vendor browsing agent or a database lookup, when volume is modest and quality matters.
So: this chain for the structured record on every row; a subagent per account for the analysis on a modest set; a vendor agent only when you are buying scale. Either way the output contract and the claim gates still bind — a subagent's reading of a pricing page is a claim like any other and needs the URL and the quoted text.
Rejected alternatives
- A metered AI browse per row. It works, but it is credits per row for a job that a free HEAD plus a free GET plus a fraction of a cent of model does at 90%.
- Entity-search APIs. You already know the entity and the URL. Nothing to search for.
- A scraping marketplace actor. Expensive, and none beats a plain GET on a public marketing page.
- Search caches and third-party pricing aggregators. Stale by construction, and a stale price in an email is worse than no price.
4. Verification
VERDICT: PASS 8/10 (80% usable) | claims 8/8 (100% correct) | multi-candidate HEAD → plain GET + price-dense window → model JSON extractor | ~$0.50/1k domains.
This is a claim-bearing playbook: 70% usable AND 90% claim correctness. unknown is a failed
fetch, not a finding, and a correct "this company does not publish pricing" is a correct outcome
that leaves the denominator entirely.
Scope note: when the record is used only to route or filter, it is internal work and the plain 70% bar applies. The moment a campaign turns the copy line on, or quotes a price, the 90% bar binds.
⚠️ The earlier 9/10 is withdrawn, and you should know why
The first run was contaminated: three of the four few-shot examples in the prompt were verbatim the scraped text and the exact expected output of three of the ten graded rows — and the prompt had been rewritten after seeing first-pass failures on those same rows.
30% of the test had its answers in the prompt.
The 8/10 above is a clean regrade: invented example companies, and 10 domains that appear nowhere in the prompt. The clean number is lower, which is what you should expect whenever a contaminated number gets re-measured.
Rule for anyone editing a prompt in this library: a few-shot example may never be text from a domain that appears in the test set, and if you tune the prompt against failures on specific rows, you must regrade on different rows.
5. Clay implementation
clay-table.md— the column build.clay-workflow.md— the CLI-buildable version.
The pricing cache
This is one of very few playbooks that earns a cache, because whether a company publishes self-serve pricing is a property of the domain rather than of the campaign, so one client's answer is reusable by the next. A row costs a multi-candidate HEAD sweep, a full page GET against the prospect's own server, and a long model grade (~3,850 prompt tokens).
Three conditions on any reused record, none optional:
- Store
last_checkedand the resolvedpricing_urlwith the record. A cached record with no date is not reusable, it is a rumour. - Re-fetch before a price goes into copy. The filter fields age reasonably; a dollar figure does not.
- This covers records you extracted from live pages. It is not permission to read pricing from a search cache or an aggregator.
Two things that sink cache designs:
- Ship the write-back in the same pass. A cache with no writer can never hit.
- Pick a window short enough to stay honest. A stale "no public pricing" verdict on a company that has since launched self-serve is a confident wrong answer in a live email — worse than the fetch it saved. Do not reuse a long email-cache window without arguing for it.
6. Locked prompt
The prompt still produces pricing_line, and that is deliberate. The default run drops the field
at the output boundary, rather than editing the prompt to stop asking for it. Editing the static
prefix would invalidate the clean regrade this verdict rests on, cool the prompt cache, and force a
full re-test to save a handful of output tokens. Keep the prompt byte-identical; drop the field in
code.
Params: max_completion_tokens=6000, reasoning_effort="low", JSON response format, no
temperature, flex tier for batch.
You are a B2B pricing-page analyst. You will be given the visible text of one company's pricing page, scraped from their website.
Your job is to decide whether this company publishes real, self-serve pricing, and to extract the plan structure. This is used to FILTER a prospect list, so a wrong "yes" is worse than an honest "no".
Return JSON only, with exactly these keys:
{"pricing_public": true|false, "pricing_model": "per_seat"|"usage"|"flat"|"credits"|"quote_only"|"unknown", "plans": [{"name": "...", "price": "...", "period": "month"|"year"|"one_time"|"unknown"}], "lowest_paid_price": "...", "has_free_tier": true|false, "enterprise_quote_only": true|false, "feature_diff_axis": "...", "pricing_line": "...", "confidence": "high"|"low"}
Rules:
- pricing_public is true ONLY if at least one plan shows a real currency amount on the page. A page that says "Contact sales" or "Request a quote" for every plan is pricing_public false with pricing_model "quote_only".
- Never invent a number. If a price is not in the input text, leave it out. Do not estimate.
- plans lists at most 6 plans, in the order shown on the page. Free tiers count as a plan with price "$0".
- Attach each price to the plan name it sits directly under on the page. Never shift a price up or down to the neighbouring plan. If you cannot tell which plan a number belongs to, leave that plan's price as "" instead of guessing.
- When a plan lists several seat types (full seat, dev seat, collaborator seat), use the FULL seat price for that plan and ignore the cheaper seat types.
- Take a price only from the plan card it belongs to. Ignore numbers that appear in body copy, footnotes or FAQs, such as a minimum invoice spend, an annual contract minimum, or a savings claim. If a plan says "Contact us" or "Talk to sales", its price is "" even when a large number appears in a sentence nearby.
- When a page shows a struck-through list price next to a limited-time promotional price ("50% off for 3 months"), use the LIST price, not the promotional one. A promo expires and would date the email.
- If the ONLY price shown for a plan is an introductory or limited-time rate ("$13/mo for 12 months", "first year", "then, starts at"), report that price, set confidence to "low", and leave pricing_line empty. An intro rate quoted as if it were the list price is a wrong claim in the prospect's inbox.
- lowest_paid_price is the cheapest list price a customer can actually pay, excluding free tiers and excluding promotions. Empty string if none is shown.
- feature_diff_axis is the single thing that changes between tiers, in 6 words or fewer, for example "seats and automation limits" or "monthly lead credits".
- pricing_line must read grammatically inside this sentence: "Noticed <pricing_line>." Write ONLY the part that replaces <pricing_line>. Do not write the word "Noticed" yourself, do not repeat the sentence frame, start with a lowercase letter, and do not end with a period. Keep it at a 5th-grade reading level.
- pricing_line names something only this company would recognise, such as the plan name plus its price. "pricing starts at $0 for the free plan" is too generic to use, so prefer the cheapest PAID plan.
- Only state a billing period in pricing_line if the page states one. If the page prices per unit with no period, say the unit, not "per month".
- No em dashes anywhere in your output. Use a comma or split the sentence.
- If the input text is a cookie banner, a login wall, a navigation shell, or is otherwise too thin to judge, return pricing_public false, pricing_model "unknown", confidence "low", and "" for pricing_line. Do not guess from the company name.
Examples (these four pages are invented, from companies that do not exist, so that no
example can ever hand you the answer to a page you are actually being asked to read):
Input: Northwind Ledger Pricing Starter $0 Free for small teams 3 workspaces Standard $7 per user/month Team $23 per user/month Enterprise Talk to us
Output: {"pricing_public": true, "pricing_model": "per_seat", "plans": [{"name": "Starter", "price": "$0", "period": "month"}, {"name": "Standard", "price": "$7", "period": "month"}, {"name": "Team", "price": "$23", "period": "month"}, {"name": "Enterprise", "price": "", "period": "unknown"}], "lowest_paid_price": "$7", "has_free_tier": true, "enterprise_quote_only": true, "feature_diff_axis": "seats and workspace limits", "pricing_line": "your Standard plan runs $7 a user each month", "confidence": "high"}
Input: Halden Systems Pricing Every deployment is scoped with our team. Book a walkthrough Request a quote Talk to sales
Output: {"pricing_public": false, "pricing_model": "quote_only", "plans": [], "lowest_paid_price": "", "has_free_tier": false, "enterprise_quote_only": true, "feature_diff_axis": "", "pricing_line": "", "confidence": "high"}
Input: Crate Studio Plans Sketch Includes: unlimited boards Studio Monthly Annual Full seat $29 /mo Builder seat $14 /mo Viewer seat $4 /mo Atelier Billed annually Full seat $68 /mo Builder seat $31 /mo Enterprise Full seat $115 /mo
Output: {"pricing_public": true, "pricing_model": "per_seat", "plans": [{"name": "Sketch", "price": "$0", "period": "month"}, {"name": "Studio", "price": "$29", "period": "month"}, {"name": "Atelier", "price": "$68", "period": "month"}, {"name": "Enterprise", "price": "$115", "period": "month"}], "lowest_paid_price": "$29", "has_free_tier": true, "enterprise_quote_only": false, "feature_diff_axis": "seat type and board limits", "pricing_line": "your Studio plan is $29 a seat each month", "confidence": "high"}
Input: Ferrous Growth Run every campaign in one place 30% off for the first year* $62 $43.40 USD per workspace / month, when paying monthly $54 USD per workspace / month, when paying annually
Output: {"pricing_public": true, "pricing_model": "flat", "plans": [{"name": "Growth", "price": "$62", "period": "month"}], "lowest_paid_price": "$62", "has_free_tier": false, "enterprise_quote_only": false, "feature_diff_axis": "billing period", "pricing_line": "your Growth plan lists at $62 a workspace each month", "confidence": "high"}
PER-ROW DATA (appended last)
PAGE DATA
Domain: <domain>
Pricing URL: <pricing url>
Visible page text:
<text>
Verifier pass: not needed for the price, and here is the better check
The model's input is the company's own live page, fetched seconds earlier, so there is no stale training data to hallucinate from. The check that actually catches fabrication is a string search:
Assert that every price and every plan name the model reports appears verbatim in the input text.
That assertion is free, runs locally, and is what graded all 10 rows. Implement it as a formula or a code assertion, not a second model call. A second model call is warranted if you ever feed this playbook cached or aggregated pricing instead of a live fetch.
Truncation guard: finish_reason=length means retry, never abstain. Measured on this exact
prompt at a 3,000-token cap: empty content on 3 of 8 rows.
⚠️ This prompt is not locked, and it is further from locked than it looks. The clean regrade exposed two new failure modes (a minimum-invoice figure captured as a plan price; an unstruck 12-month intro rate reported as list price). Each got a new rule above, and neither rule has been tested since it was added.
7. Edge cases and failure modes
| Symptom | Cause | Fix |
|---|---|---|
| A quote-only company is reported as having public pricing | {domain}/pricing returns HTTP 200 and redirects to a 404 page. Measured on a real enterprise domain resolving 200 to /404 | Require the final resolved path to still match /pricing|plans|price/. Status 200 alone proves nothing. A guard that fails in the wrong direction records every quote-only company as having public pricing — a false claim on every one |
A page with real prices is reported quote_only with high confidence | The text was truncated from the front and the price table was past the cutoff. Measured: one page extracts to 354,059 chars and the first 14,000 contain zero prices | Use a price-dense sliding window, never text.slice(0, N) |
| Rows come back empty with no error | finish_reason: "length", budget spent on reasoning. 3 of 8 rows at a 3,000-token cap | reasoning_effort: "low", cap 6000, retry on length. Never record it as an abstain |
| You paid for a render and got the same empty JS shell | A JS page may only fill in after paint | Keep a wait as the default but treat it as unproven. An A/B on one real page: no wait = 43 currency tokens in 3.4s; wait=8000 = 43 currency tokens in 23.5s, and the extra characters were the cookie banner. If you are burning credits with nothing to show, A/B it on your own pages |
| A "Contact us" plan comes back with a big price attached | A number in body copy or a footnote sits near the plan table. Measured: a page publishing "available on the Standard or Teams plan for $5,000 and Enterprise plan for $15,000" had $15,000 attached to an Enterprise tier the same page marks "Contact us" | The prompt now forbids it. That rule is untested — spot-check any plan whose price is an order of magnitude above the rest |
| The email quotes an intro rate as the list price | A promo with no struck-through comparison, so a struck-through rule never fires. Measured: a page showing "$13.00/mo for 12 months ... Then, starts at" where the list price never reaches the text layer at all | Report the intro price, set confidence: "low", leave the copy line empty. Untested. Treat low rows as filter-only |
plans[].price contains period text ("$20 per user / month") | The model folded the period into the price field. 2 of 4 plans on one domain | Cosmetic — lowest_paid_price stayed clean, so the filter is unaffected. Normalize with a regex before pushing to copy |
lowest_paid_price names a number in no plan in the same record | Annual and monthly rates both appear; the model picked annual for one field and monthly for the other. Both true, both on the page | Not a fabrication, but do not use both numbers in the same sentence |
| A pricing page exists but the origin refuses HEAD | The GET fallback's Range request gets 206, not 200 | Accept any 2xx |
| A CDN-fronted site looks like it has no pricing page | A default runtime User-Agent got 403'd | Send a browser User-Agent on every request. Never validate this path with curl — curl is not the production path |
The copy line renders Noticed noticed your Pro plan... | The prompt stated the sentence frame and the model echoed the frame word | The prompt explicitly forbids writing "Noticed". Re-check after any edit |
| Plan names and prices off by one tier | Multi-seat-type cards confuse the alignment | The prompt pins price to the plan name it sits under. Verify with the free string-search assertion |
plans lists product SKUs instead of tiers | Very large multi-product pricing pages have no single tier table | Accept it. pricing_public and pricing_model are still correct, which is what the filter uses. Do not put those plans into copy |
| Some page never yields prices no matter what | The plan data arrives in a later XHR, not in the HTML | Accept the miss, or capture the vendor's own pricing XHR once with a real browser and hit that endpoint directly |
| Transient connection resets | Normal network noise | Retry with backoff. A network failure is never a verdict, in either direction. A fetcher with no retries is worse than it looks: an empty body fails the richness gate and spends a metered render credit on what was only a blip |
| The prompt cache stops paying and the estimate drifts | Any edit to the static prefix cools the cache, and small parallel batches never warm it. Measured: 6 of 8 calls cached on one run, 0 of 8 after the prefix was edited | Budget the uncached number. Re-read the cached-token count after every prompt edit instead of carrying the old figure forward |
Hard rules
- Bones by default, no pricing line. Do not produce, push, or build a column for the copy clause unless the brief asks.
- Every price and plan name must appear verbatim in the fetched page text. Assert it in code. A price a prospect cannot find on their own site destroys the message — and writing "Base" where the page says "Basic" is the same failure. Never grade a batch by eye; grep every reported name and price back out of the captured text.
- A true negative is only confirmed by a second, independently located page with real readable text. A second source that comes back as a 26-character JS shell confirms nothing in either direction.
unknownnever removes a row.- Fetch politely: concurrency 2 per origin. These are the websites of companies you want to sell to.