Playbook google site search skill

Playbook: Google site: Keyword Filter

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: a campaign needs to know whether each company's own website talks about a specific thing — a certification, a technology, a service line, a location, a program name — and no structured API sells that field.

Do not use when: you need to find companies you do not have yet; when the thing you want is a page rather than a phrase (playbook-pricing-page, playbook-case-study-page); or when a real filterable field already exists (funding stage, headcount, technology installed) — filter at the source and skip this entirely.

One-line output: site_keyword_line = "you hold SOC 2 Type II certification", rendering as Noticed you hold SOC 2 Type II certification.

1. Trigger and scope

You have a list of domains. Someone asks "which of these mention X on their website?"

It does two jobs at once, and you should know which one you are buying. As a filter it tells you which rows to keep — that half is reliable. As a personalization signal it gives you an evidence URL and a line — that half needs the confidence gate in §2, because a page can contain your keyword for reasons that have nothing to do with the company.

It does not prove a company lacks something. The index is incomplete and its results are not even stable between two identical calls a minute apart (measured). An empty result means "not found in the index today", never "this company does not do X".

The rule that defines this playbook: one keyword, one query

site:acme.com "chess". Never site:acme.com ("chess" OR "checkers" OR "board games").

Measured:

Query	Precision
site:n8n.io "SOC 2"	10/10 results contained the phrase
the same domain with a 7-term OR chain	4/10
site:uschess.org "Brooklyn"	8/10 on target
the same with an OR chain	2/10, and it drifted to a chess house in New Orleans

The chain also destroys attribution: one response for seven keywords means you cannot tell which one matched.

2. Output contract

Inputs required per row

Field	Type	Required?
domain (bare, lowercase, no scheme, no www)	string	yes
keyword	string, the exact phrase	yes
company_name	string	no, improves readability

The keyword is a campaign-level constant, not a per-row value. One campaign, one keyword. Three keywords means three columns and three queries per row, each with its own answerable yes or no.

Output fields

Field	Type	Example	Null?
site_keyword_hit	"yes" or ""	yes	yes
site_keyword_line	string	you hold SOC 2 Type II certification	yes
site_keyword_url	URL	https://linear.app/security	yes
site_keyword_status	has / planned / discusses / none	has	no
confidence	high / low	high	no
campaign_usable	boolean	true	no
site_keyword_sentence	the whole pre-rendered sentence with its period and one trailing space	Noticed you hold SOC 2 Type II certification.	yes

campaign_usable is the gate, and it is deliberately stricter than site_keyword_hit. True only when status == "has" AND confidence == "high".

⚠️ site_keyword_line is emitted ONLY when campaign_usable is true

It is "" otherwise — including on rows where site_keyword_hit == "yes". This is not a downstream convention an operator is trusted to remember. It must be computed that way, and campaign_usable must be computed first.

Why this is written this hard. In the graded test the judge returned, for one domain:

{"mentions": true, "status": "discusses", "confidence": "low",
 "line": "you passed your SOC 2 audit on the first try"}

The evidence URL was a prompt-template marketing page on that company's own site — the sentence had been scraped out of an example cold email printed inside their own content. That company had passed no such audit. site_keyword_hit was yes, so any pipeline that ships the line whenever there is a hit sends a fabricated compliance claim about the prospect's own company, to that company.

A row that is a hit but not usable is still a perfectly good filter answer. Keep the hit flag for segmentation. Just never let it drive copy.

Coverage expectation

9/10 rows usable (90%) in the graded test: six genuine hits, four abstains (three correct).

Read that carefully before making the variable mandatory. On a real list roughly half to two thirds of rows will abstain, because most companies do not mention most keywords. The 90% is the accuracy of the answer, not the fill rate.

And the 90% is retrieval accuracy, not copy quality. Among rows that cleared the gate, roughly 1 in 6 carried a grammar or brand-casing defect a human would have to fix by hand. Measured: one row shipped campaign_usable=true with "you have SOC 2 Type 2 attestation" — missing an article. Another shipped "linear undergoes regular SOC 2 audits" with the brand lowercased, because the model itself wrote it lowercase, and a sanitizer's brand guard only prevents the sanitizer from lowercasing — it does not re-capitalize what arrived lowercase.

Neither defect is caught by anything downstream. That is why §7 makes a 20-sample human grammar read a required pre-launch step.

Downstream gate: the half to two thirds that abstain

⚠️ Spintax cannot do this job. It picks a variant at random and cannot branch on whether a variable is empty. {Noticed {{site_keyword_line}}.|Quick one for you.} still renders Noticed . half the time on abstain rows.

Mechanism A (default): pre-render the whole sentence, ship a blank on abstain.

That is what site_keyword_sentence is for. It contains the entire sentence including its period and one trailing space, or "". Put it in the body by itself, with nothing around it:

Hi {{first_name}},

{{site_keyword_sentence}}{{value_prop_line}}

On a usable row: Noticed linear undergoes regular SOC 2 audits. Most teams we work with... On an abstain row: Most teams we work with... — no gap, no stray period, no leading space. Nothing is left to an operator's memory, because the emptiness is baked into the value.

Mechanism B: segment the list. Filter on campaign_usable == true at upload and never let an abstain row into the campaign. Then the variable is mandatory and Mechanism A is unnecessary. Cleaner copy, smaller list.

At a 50-66% abstain rate, Mechanism B is usually the right call — running the covered rows as their own campaign beats papering over two thirds of the list with blanks.

3. Source chain (cost-tagged)

#	Source	Cost	Call	Result
1	SERP API + literal-phrase filter + model judge	CHEAP	site:{domain} "{keyword}", limit 10, then the §6 judge	measured 90% usable, 6/10 rows were hits
2	A semantic entity-search API	METERED	for a single high-value row only	untested
3	Direct fetch of guessed paths (/security, /trust) through a rendering proxy	METERED	grep the literal phrase	untested. Only when you specifically suspect the page exists but is not indexed

No cache. Recompute — it is cheap. At roughly $0.59 per 1,000 rows blended, a repeat check on a domain another campaign already ran is worth less than a tenth of a cent. If a list has real internal overlap, dedup on (domain, keyword) before the run.

Cost, and a correction worth internalizing

Leg	Unit	Per 1,000 rows
SERP call	~$0.0005 per request	$0.50
Model judge	1,335 in + 53 out per row	$0.09
Blended		$0.59

An earlier draft put this at $0.09/1k by counting only the model leg, on the assumption that the SERP plan was flat-rate with no per-call charge. That assumption was wrong. Still cheap, but budget against $0.59, not $0.09. When a cost line looks suspiciously low, check whether someone counted only the leg they could see.

Rejected alternatives

• OR-chain queries. See §1. Precision fell from 100% to 40%, and 80% to 20%, on two different cases.
• Raw keyword matching without the judge. One domain returned 7 on-domain results all literally containing the phrase, and every one was a marketing or prompt-template page, not a claim about the company. A naive filter marks that company a match.
• Google's official Custom Search JSON API. 100 free queries per day then metered, with a 10k/day hard ceiling. Useless at list scale.
• A browsing agent per row. It works, but it is credits per row for something one search plus one cheap model call does for a tenth of a cent. Keep browsing agents for jobs that need a browser.
• Bare site:domain with no keyword. Returns whatever the engine feels like ranking and answers no question.

4. Verification

VERDICT: PASS 9/10 (90%) | SERP site: query → literal-phrase filter → model judge at minimal reasoning effort | p50 1.7s/row | ~$0.59/1k blended.

Tested on 10 real B2B software domains with the keyword SOC 2, including 2 true negatives and 3 hard rows. Zero errors, zero rate limits across ~54 calls at concurrency 2. The one failing row was a copy-grammar failure, not a retrieval failure.

A post-review fix worth knowing about. A verifier pass found the shipped script set site_keyword_line on every hit, not only on gate-passing rows — so the fabricated compliance claim in §2 would have reached the sequencer despite the documented gate. The gate was documented correctly and implemented wrongly. That is the failure mode to watch for in your own build: a rule that lives only in prose.

⚠️ This verdict covers the script path. It does not cover the Clay implementation.

5. Clay implementation

Step 0, before you build anything: verify the search by hand

Run the exact query for 3 to 5 real domains from your actual list and read the results with your own eyes before you build a single column. This is not optional.

A query that looks obviously right can return a completely different company: site:patagonia.com "SOC 2" returned 10 results, zero of which contained the phrase — all matching a company whose name contains "Soc Trang". Ten minutes here saves a rebuilt table.

You are checking four things. Any one failing means fix the keyword before you scale:

1. Are the results actually on the domain you asked for?
2. Does the snippet literally contain your keyword, or did the engine fuzzy-match?
3. When the keyword is there, is it the company talking about itself — or a blog roundup, a customer logo wall, a job post, a forum thread?
4. Does the domain you expected to miss actually come back empty? If everything hits, your keyword is too generic to filter on.

Then:

• clay-table.md — the 11-column build, including a dependency cycle that will otherwise stop your table from building at all.
• clay-workflow.md — the CLI-buildable version.

6. Locked prompt

Model: a nano-class reasoning model — at 1,335 input plus 53 output tokens per row it is cheaper than a mini-class model here (~$0.088/1k vs ~$0.23/1k), so nano wins inside Clay too, which is the opposite of most playbooks in this library.

Params: max_completion_tokens=1200, reasoning_effort="minimal", JSON response format, no temperature.

Minimal reasoning effort is load-bearing, not a tuning preference. Without it, the model burned past 1,800 completion tokens on reasoning and returned finish_reason=length with empty content on 2 of 10 rows. With it: measured 0 reasoning tokens, and judge latency fell from 17.6s to 1.1s.

You judge whether a company's own website really talks about a given keyword, and you write one short line of outreach copy about it.

You will be given: a company domain, a keyword, and up to 5 Google results returned by the query site:<domain> "<keyword>".

Return JSON only, exactly these keys:
{"mentions": true|false, "status": "has"|"planned"|"discusses"|"none", "evidence_url": "", "line": "", "confidence": "high"|"low"}

How to decide "mentions":
- true only when a result is a page the COMPANY published about ITSELF and the keyword describes that company, its product, its service, its customers, or its own credential.
- false when the keyword only appears because the page is a directory listing, an integration or partner page about a DIFFERENT company, a user forum post, a job board scrape, a customer logo wall, a press roundup, or a generic blog post not about this company.
- false when no result is from the company's own domain or a subdomain of it.
- false when the only match is an unrelated word that happens to contain the keyword letters.

How to set "status":
- "has" when the page shows the company already does or holds the thing.
- "planned" when the page shows it is on a roadmap, in progress, requested, or coming soon.
- "discusses" when the company writes about the topic but the page does not show it holds or plans the thing itself.
- "none" when mentions is false.

Rules for "line":
- It must complete this sentence with correct grammar: "Noticed <line>."
- NEVER start the line with "Noticed", "noticed", "I noticed", "that", or "you have". The word "Noticed" is added automatically in front of your line. Starting with it produces "Noticed Noticed ..." which is broken copy.
- Write what you noticed about THEM, as a fact about the company. Never describe the web page itself. Do not write "is mentioned on", "the page says", "listed on your site", "in your security documents".
- Start with a lowercase letter UNLESS the first word is a proper noun, a brand name, or an acronym. Keep acronyms and brand names in their correct capitalization everywhere (SOC 2, HIPAA, HubSpot, ISO 9001).
- No trailing period. No em dashes. No quotes inside it.
- 5th-grade reading level. Under 90 characters.
- State only what the snippet proves. Never invent a fact, a date, or a number. If status is "planned", say it is planned.
- If mentions is false, return "" for line and "" for evidence_url.

Rules for "confidence":
- "high" when the snippet itself contains the keyword and the page is clearly the company's own page.
- "low" when you inferred it, or the page could belong to someone else.

Examples:
Input: domain=linear.app keyword=SOC 2 results=[{"title":"Security","url":"https://linear.app/security","snippet":"Linear undergoes regular Service Organization Controls audits (SOC 2 Type II)."}]
Output: {"mentions": true, "status": "has", "evidence_url": "https://linear.app/security", "line": "you run SOC 2 Type II audits", "confidence": "high"}
Input: domain=smallco.io keyword=SOC 2 results=[{"title":"Roadmap","url":"https://feedback.smallco.io/roadmap","snippet":"SOC 2 security certification - planned"}]
Output: {"mentions": true, "status": "planned", "evidence_url": "https://feedback.smallco.io/roadmap", "line": "SOC 2 is still sitting on your public roadmap", "confidence": "high"}
Input: domain=midco.com keyword=SOC 2 results=[{"title":"Trust","url":"https://trust.midco.com/","snippet":"We are in the middle of our SOC 2 Type 2 audit window."}]
Output: {"mentions": true, "status": "planned", "evidence_url": "https://trust.midco.com/", "line": "you are partway through your SOC 2 Type 2 audit", "confidence": "high"}
Input: domain=acme.com keyword=chess results=[{"title":"Best chess apps 2026","url":"https://acme.com/blog/roundup","snippet":"We ranked 20 chess apps."}]
Output: {"mentions": false, "status": "none", "evidence_url": "", "line": "", "confidence": "low"}
Input: domain=acme.com keyword=SOC 2 results=[]
Output: {"mentions": false, "status": "none", "evidence_url": "", "line": "", "confidence": "high"}

PER-ROW DATA FOLLOWS.

The static prefix measures about 1,030 tokens — just over the 1,024-token caching minimum, so at volume it caches at roughly a 90% discount. Interpolating row data into it kills that discount.

Per-row data goes last, as the user message:

domain=<domain>
keyword=<keyword>
results=[{"title":"...","url":"...","snippet":"..."}]

Verifier pass: not needed. The judge never sees the open web, only snippets returned for this exact domain, and the prompt forbids stating anything the snippet does not prove. The hallucination surface is a claim drifting from its snippet, which the deterministic guards catch more cheaply: the literal-phrase filter proves the keyword is in the text, and the evidence URL lets a human check in one click.

⚠️ The one thing to watch is a stale snippet — a search cache can be months old. If your keyword is time-sensitive (a funding round, a launch date, a "coming soon"), this is the wrong playbook.

Truncation guard: finish_reason=length means retry, never abstain. After 3 failed attempts, record the row as an ERROR and exclude it — never as an abstain.

7. Edge cases and failure modes

Symptom	Cause	Fix
Results returned but none contain the keyword	Search engines do not strictly honor quoted phrases on site: queries. One query returned 10 results matching an unrelated company whose name shared the letters	A deterministic literal-phrase filter over title plus snippet before the model runs. Never trust the raw hit count
The line quotes a stale claim that contradicts the company's current page	site: includes subdomains, and community. / forum. subdomains are users talking, not the company. A stale forum answer outranked the company's own current security page	Drop community. forum. forums. discuss. answers. feedback. ideas. status. help. hosts before judging
Every row on a domain "matches" but the claims are nonsense	The keyword appears in blog roundups, prompt templates, integration directories, customer logo walls	The status + confidence gate. Only has + high reaches copy
Copy renders Noticed Noticed you hold SOC 2.	The model prepends the sentence frame. Happened on 3 of 10 rows despite an explicit prompt rule	A deterministic sanitizer strips a leading "Noticed"/"I noticed"/"that". A regex, not a prompt rule
Copy renders Noticed .	The model returned mentions: true with status: "none" and an empty line	A row is a hit only when mentions === true AND status !== "none" AND the sanitized line is non-empty
Copy renders Noticed . on half the list	The body used Noticed {{site_keyword_line}}. and the variable is correctly blank on every abstain row	Use the pre-rendered sentence field alone, or segment on campaign_usable. Spintax cannot branch on emptiness
A junk row for the domain "domain"	A CSV header row processed as data — burning a real search call	Skip a first line whose first field is literally domain
Brand name lowercased	The sanitizer applied lowercase-first blindly	Skip lowercasing when the first word is an all-caps acronym, carries an internal capital, or matches the domain root. ⚠️ This does not fix a name the model itself wrote lowercase
The same query returns 7 results, then 0, then 7	Search results are genuinely non-deterministic between calls. Observed on three domains	Retry once before recording a hard negative. Never treat one empty response as proof of absence
Empty content, finish_reason=length	Reasoning overrun	Minimal reasoning effort plus retry. Exhausted retries is an ERROR row, never an abstain
The HTTP column 400s or 431s	Raw quotes and colons in the URL, or oversized headers	Percent-encode the whole query string. Set automapInputs: false on workflow nodes
The evidence URL does not contain the quoted phrase	The judge quoted one result and cited another	If the URL is load-bearing, verify the phrase is on that specific page
Fetching the evidence page returns nothing	Trust centers and feedback boards are JS-rendered	Verify against a first-party page on the main domain, or use a real browser

Required pre-launch step: the 20-sample grammar read

Before any campaign using this variable goes live, a human reads 20 rendered samples. Not optional, and not a spot check you skip when the batch "looks fine".

The gate catches false claims. Nothing in this playbook catches bad grammar, and bad grammar is measurably common — roughly 1 in 6 rows that clear the gate.

1. Pull 20 random rows where campaign_usable == true.
2. Render each into the actual frame: Noticed {{site_keyword_line}}.
3. Read all 20 out loud. Mark any that need an edit of any size.
4. Grep the rendered output for em dashes, double periods, a leading "Noticed Noticed", and a stray "Noticed ." with nothing between.
5. If more than 1 of 20 needs an edit, do not launch. Fix the prompt, re-run, read a fresh 20. If a single brand name keeps arriving lowercased, add it to the sanitizer's proper-noun guard rather than hand-editing rows.

Note step 4 lints rendered output, not raw field values. A field that looks fine in a table can render broken in a frame.

Hard rules

• One keyword, one query. Never an OR chain. This is the defining rule of this playbook.
• Manually verify your searches before scaling. Step 0 of §5.
• The hit count is never the filter, an error is never an abstain, and the line is blank unless campaign_usable is true. A 429, a 5xx, an empty judge cell or an exhausted retry is an ERROR row: re-run it. A hit is a filter answer, not permission to ship copy.
• Check your SERP provider's response shape. A "light" endpoint often returns a flat array while the full endpoint nests results one level deeper — and parsing one like the other silently returns zero rows rather than erroring.