What would prove us wrong?

The belief under test

<the belief the plan depends on, in one sentence>

Every other playbook here starts from a question the instruments were shaped for. This one starts from an answer you already hold and tries to kill it. The procedure does not change with the subject — the same gates work on a demand claim, on a rival's strategy, on a traffic drop, and on a disagreement between two people who both have data. What changes is where you aim, not how you decide.

It is also the cheapest run on this list, and that is the argument rather than a concession. Most of the work is free, and the paid part is short because most calls you could make cannot change the answer.

THE EVIDENCE LADDER. Rank every signal by what it would cost someone to fake. Prefer signals higher up, and never let a lower rung outvote a higher one:

1. Money someone is paying a competitor, or sustaining ad spend
2. Salary a company is paying humans to do this job
3. Search commercial-intent queries with real, repeated volume
4. Time people build workarounds — spreadsheets, templates, macros
5. Attention views, engagement, followers, upvotes
6. Words someone said they would like it. Never sufficient on its own.

Rungs 1 and 2 cost money to fake. Rung 6 costs nothing, which is why so much of it exists.

FINDING URLS. Several tools take a URL and cannot search for one — social_linkedin_posts, social_reddit_posts, social_reddit_comments, company_jobs, company_person, web_fetch. When you need a page, find its URL with YOUR OWN web search first, using a site: filter. That costs nothing and is exact. web_search does the same job for a billed call and is the fallback when you have no search of your own. Never construct a URL from a company name: a wrong URL comes back as a dead page, which reads as an absence of evidence rather than as your mistake.

BATCHING. The keyword tools bill once per call however many terms you send — keyword_volume takes 1,000, keyword_intent 1,000, keyword_overview 700, keyword_ideas 200. Pool your terms and send them in one call rather than looping. A batch whose rows exceed one response comes back trimmed and says how many rows it held versus returned: size the next batch from that figure and send the terms you did not get. A term missing from a trimmed response was withheld, NOT measured at zero. Scoring it as zero is the single most common way one of these runs reaches a confident wrong answer.

Steps 1 to 4 cost nothing. Many runs should end inside them.

The expensive part of research is the calls. The cheapest way to make a call worth its price is to know, before it returns, which result would change your mind.

1. Write the belief as an explanation, not a prediction

"Demand is growing" is a prediction. "Demand is growing because the rules changed in March, so buyers who used to handle this in a spreadsheet now have to file it" is an explanation. Only the second tells you where to point an instrument, because only the second says what else would have to be true.

Write the mechanism. If you cannot write one, you do not yet have something to test — you have a number you expect to see, and any number can be explained after it arrives.

2. Try to save it, in writing. This gate kills the most runs.

Suppose the data comes back against you. Which clause of your explanation would you adjust to keep the belief alive? Write that clause down now.

An explanation you can rescue from any result explains nothing, because it forbids nothing. If you found such a clause you have two choices and only two: fix it — commit in writing to the version you will not move — or record that the belief is not yet testable and stop.

Do this before the calls. Afterwards it is indistinguishable from reasoning.

3. Name the observation that would refute it, and check that it exists

State one result, from one named tool, at one number, that you would accept as a refutation. Not evidence you would weigh. A refutation.

  TOOL       the call you would make

  ARGUMENT   what you would send it

  RESULT     the shape of answer that kills the belief

  THRESHOLD  the number, fixed now

If no public observation could bear on the belief either way, it does not touch anything this API can reach. No amount of spending fixes that. Say so and stop.

KILL CRITERION 1, at zero calls. No tool, no argument, no threshold — no run. Report it as the finding. A belief that survives because it was never exposed is the most expensive thing on this list, and it costs nothing to catch.

4. Write the rivals. At least two, and make one of them boring.

You cannot read an explanation off a dataset. Every number is consistent with several stories, and the one you notice is the one you walked in with. So bring the others deliberately.

Write at least two rival explanations for the same evidence. The boring one is usually right and is almost never the one a team arrives with. Four account for most of it:

  THE INDEX MOVED         coverage or measurement changed, not the subject

  THE CATEGORY MOVED      everyone rose or fell together; you did nothing

  THE INSTRUMENT IS BLIND thin coverage here, not an empty market

  NOTHING MOVED           you are reading a level and calling it a trend

A run with one candidate can only confirm. Confirmation costs the same 20¢ as refutation and is worth less.

Where do rivals come from when you have none? From people describing the same thing differently. Find the thread with your own search, or web_search, then social_reddit_posts on the post itself — it takes the post's own URL and refuses a subreddit front page. Read it for candidate mechanisms, not for a verdict. Anecdote sits at the bottom of the ladder and its job here is to generate explanations, never to choose between them.

Steps 5 to 9 are the run itself. This is where the money goes.

5. Derive the crucial test — and pay for nothing else

For each pair of belief and rival, write what each predicts for a specific call. Then spend only where the two predictions differ.

An observation both explanations predict cannot separate them. It will come back, it will feel like progress, and it will move nothing. That is the whole budget rule here.

A belief about demand — "people want this", "this is growing".

Not keyword_volume on its own: a large number is predicted by real demand, by a fad, and by having picked terms broader than your product. Instead one keyword_history call — up to 700 terms for a single charge — carrying your terms AND a control set from an adjacent category you believe is unaffected. Your terms rise while the controls do not: the subject moved. Both rise: the season or the index moved. Then keyword_intent on the same list, which separates people who want to buy from people who are curious. The two have the same volume and different worth.

A belief about a rival — "they are betting on this", "their strategy is that".

Not their website. Everything a company publishes is written to be read by you, so web_fetch on a page returns what they claim, which no rival explanation disputes. Instead ads_creatives with target=<their domain>, and read first_shown and last_shown on every row. A creative running 90+ days survived their own performance review; nobody sustains a losing ad for a quarter. Then company_jobs on one or two live listings, found with your own search. A message they pay to keep running and a role they pay a salary for are both claims that cost money to make. A homepage is free.

A belief about a move — "we lost ground", "the market shifted", "they took our traffic".

Not a solo call on your own domain. Instead domain_traffic_history with targets=<you>,<rival>,<rival> and date_from at least 24 months back. This is the crucial test in its cleanest form: one argument different, and the two explanations predict different shapes. A month where every domain in the batch falls together and recovers the next is the instrument, not the market — and you can only see that because you asked about more than one. Then domain_history on the subject, and read the position bands rather than count.

A belief about your position — "we are the obvious choice", "they would pick us".

Not your own list of competitors. Instead domain_serp_rivals on the terms a prospect would actually type: it takes the keywords, not your domain, and returns who shows up when the buyer asks. Then keyword_suggestions on "<brand> alternative", for yours and for theirs, which measures the population each current answer has already pushed away.

6. Run it. Four observations that look decisive and are not.

Each of these has already produced a confident wrong answer in a real run of another playbook here.

A falling count is not a decline. One subject's ranked-keyword count fell 60% over six months while it held more top-ten positions at the end than at the start. A pruned tail and a collapse both predict a falling count, so count cannot separate them. domain_history's position bands — pos_1, pos_2_3, pos_4_10 — can. The tail is most of count and almost none of the money.

A name match is not an advertiser. ads_advertisers matches loosely on the name: run on a short brand it returned four verified advertisers, none of them the subject, while the subject's own forty creatives came back from ads_creatives with target=<domain>. Go to the creatives first and use ads_advertisers only to disambiguate, checking title or domain on every row before you attribute one ad to anyone. Borrowed evidence is worse than no evidence, because it reads as evidence.

An empty result is not an empty market. Run the same call, unchanged, on a control you are certain about — a term you know has volume, a domain you know ranks. If the control comes back empty too, the instrument is blind for this query or this country and the belief is untested, not refuted. One extra call buys the difference between "there is nothing there" and "I cannot see". Make it before you report an absence.

A withheld row is not a zero. BATCHING above says it, and it belongs here too, because it is the same error wearing different clothes. A trimmed response reports how many rows it held against how many it returned. The terms you did not get were withheld, not measured at zero.

7. If it was refuted, the run finished. It did not fail.

Write down the variation you were tempted to make.

You will feel it — the clause from step 2, offering itself again now that you know which way the data went. Record it in the words you would have used. That record is the most valuable thing this run produces, because in three months somebody will propose exactly that variation, and by then there will be no memory that it was invented to save a belief the evidence had already killed.

Then stop. A refutation reached in four calls is the best outcome available here, and it is what the free gates were for.

8. If it survived, make it pay: name what else it predicts

A conclusion that accounts for exactly the data you bought is not an explanation, it is a summary of your receipts. An explanation worth keeping reaches further than the problem it was built for — it makes claims about things you were not looking for, and those claims can be checked.

So state one consequence of the surviving explanation that you have NOT checked and that the rivals do not share. Then either buy it in one call or write it as the first call of the next run.

KILL CRITERION 2. If the surviving explanation predicts nothing beyond the observations you already paid for, you did not learn a mechanism. You learned that some numbers moved together. Say that in those words, so it cannot be quoted later as the reason for anything.

9. Survived is not true. Write the expiry.

It survived this test, at this threshold, in this country, on this date, with these instruments. All five are load-bearing and all five decay. Expect your best current explanation to contain something false; that is the normal condition, not a failure of this run.

So finish by writing what would have to change for the run to be worth repeating — a rival launching, a season turning, the threshold crossed from the other side. A belief with no scheduled re-test becomes an assumption again inside a quarter, and the next run will start from it instead of testing it.

Where the instruments are blind

• Mechanism is what you are testing and what public data is worst at. You

  can measure that something moved, and with controls whether it moved for the

  subject or for everyone. WHY it moved is inference. So the strongest verdict

  usually available here is "this rival is now excluded", not "the belief is

  confirmed" — and the first is worth more anyway.

• Intent is not localized. keyword_intent takes a language and no

  country, so for a belief about one market the intent split is measured across

  places rather than in one. `keyword_volume` and `keyword_history` do take a

  location; localize with those and read intent as a shape, not as a level.

• Coverage thins outside the largest markets. An absence found far from the

  United States is much more likely to be the instrument than one found inside

  it, which is why the control call matters more the further out you go.

• Nothing here sees your own numbers. A belief about what YOUR change caused

  cannot be tested from outside. What can be tested is whether the world around

  it moved at the same time, and whether it moved for your competitors too.

The verdict

One of four, and only these four:

  REFUTED       the discriminating observation came back against the belief.

                Give the call, the number, and the threshold fixed in step 3.

  SURVIVED      it came back for the belief, and at least one rival is now

                excluded by name. Give the rival, and what excluded it.

  INCONCLUSIVE  the call was made and did not separate them after all. Name

                the one further call that would, as the rules below require.

  UNTESTABLE    nothing available here could have separated them, and gate 3

                should have caught it. Say what a different instrument would

                need to see.

The last two are different failures and the difference matters: INCONCLUSIVE is an instrument that came up short, UNTESTABLE is a question that was never exposed to one. "Supported" is not a verdict. Neither is "directionally correct". A run that cannot say which of the four it reached has not finished.

RULES THAT MAKE THE OUTPUT TRUSTWORTHY

• Cite the tool and the argument behind every number you state.
• Distinguish measured (a tool returned it) from inferred (you reasoned to it).
• State the strongest evidence AGAINST your conclusion before your conclusion.
• Absence of a signal is not evidence of absence. Name which instrument was

  blind and why.

• Coverage is public data and varies by query, country, and date. A thin

  result is a thin result, not an empty market.

• If the data does not settle it, say INCONCLUSIVE and name the one call that

  would settle it. An unspent budget and an uncertain answer at the same time

  is a failed run; so is a confident answer the data did not support.

Deliver

The belief as you wrote it before the first call, and the clause you committed not to move. The rivals. The observation you chose, and why it discriminates rather than merely relates. What it returned. The verdict, one of the three. The variation you were tempted to make. And, if it survived, the one prediction it makes that you have not yet paid to check.