How to share creative learnings: the one-page note, the hypothesis card and the monthly review

What must a creative report answer?

A creative report must answer one question: what should the team make, fund or stop next, and why does it believe that? A list of the top ten ads does not answer it, because the next ad will not be one of those ten. A learning about the buyer does. Here is an invented one: "Finance managers who sign for a whole team look for a peer with their own job title first." It tells a designer what to make. "Ad 4 won" does not.

Every reader also asks a private question before they read a number. The head of marketing asks whether the next budget meeting holds a surprise. The finance reader asks whether the spend is safe to repeat. The client asks whether you spent the money well. The designer asks whether the feedback is about their work or about the buyer. This is our reading, and it shapes the page below: the first line answers the private question.

The three public LinkedIn reporting templates we read, all from marketing analytics vendors, rank single ads, creatives or formats. One has a creative performance section that says to compare creatives within each campaign. Another lists a section for the ads with the strongest results. The third says to identify which ad creatives and formats drive the best results. We read all three in October 2026. None of them says to group ads by angle or to keep a log of learnings, and none gives a form for the next test. We take two things from them: a fixed order for the report, and the rule that every recommendation ties to a specific data point in the report.

The template below is the whole report. Each part has its own section after it. Paste it into the monthly document. In Slack, post the decision line and one picture, and link the page.

CREATIVE LEARNINGS: [account or product], [period]
Read on: [date]. The latest days may still change.
Goal and outcome: [campaign goal]. The campaign buys [outcome]. Goal metric: cost per [outcome].
Basis: [platform], [conversion rule windows or ad set attribution setting]. Numbers from different platforms are not added.
Enough data means: [results per group, set before the period]. Groups below it are marked too early.

DECISION
[One line: what we move, make or stop, and why.]

ANGLE: [name in the buyer's terms]
Ads: [ad images, or ad names]
Result: [total spend] / [total results] = [cost per result]. Ads in the group: [number]. Too early: [yes or no]
What it tells us about the buyer: [one sentence]
Verbatim line: "[exact words from the ad]"
Because [observation across ads], we believe [cause]. Next we test [concept] against [control], judged on [goal metric].
(Repeat the ANGLE block for every angle.)

OUTLIERS: [single ads that fit no group, and why they are not a pattern]
SPENT, NO RESULT: [ads]
TOO EARLY: [angles with too few results]
OPEN QUESTIONS THAT CHANGE A DECISION: [only those]

How do you group results so a reader sees the point?

Group by angle: the reason you give the buyer to act. Ads in one group can differ in format, hook and picture. They share one reason. The single ads are the proof of the group, not the subject of the report.

Follow five rules:

1. Name each angle in the buyer's terms, such as "the auditor asks", and not "Angle 3".
2. Put every ad in one group.
3. Put an ad that fits no group on the Outliers line. One ad is an outlier. A pattern needs two or more ads that share the reason and differ in other ways.
4. List ads that spent money and got no result on their own line. They have no cost per result, so a ranking by cost never shows them.
5. Mark a group as too early when it has few results (see the section on numbers).

Get the groups from the ad name. Add the angle when you make the ad, for example 2026-09_LI_peer-number_title-first_v2 (date, platform, angle, hook, version). The export then carries the group, and nobody sorts forty ads by hand before the review. Nimblet reads ad-name conventions, "contains" rules and manual tags for the same job.

The tables below use an invented product, an expense tool for finance teams of 50 to 200 people. The numbers are invented too. The campaign buys one outcome: a lead. The goal metric is cost per lead (CPL), which is spend divided by leads.

Add the spend and the leads first. Then divide once. Do not average the CPL of the ads. For the auditor angle, the average of €68.75 and €100.00 is €84.38. The true CPL is €1,700 / 22 = €77.27, because the €1,100 ad carries more of the money. In the time-back group, T2 spent €150 and brought no lead, so it has no CPL. An average of CPLs drops it without a word. The group sum, €450 / 3 = €150.00, keeps its money.

The table shows where to look. It does not show why. Ad sets differ in audience, bid and budget, so a gap between two groups can come from the audience and not from the angle. This is why each group ends in a test against a control.

How do you write a learning the team can test?

Write it as a claim about the buyer that the next test can prove wrong. Use this form: Because [observation across ads], we believe [cause]. Next we test [concept] against [control], judged on [goal metric]. A tool can draft the observation. A person signs the cause and the bet, because the client or the finance reader pays for that judgment.

Run the junior test before you keep a learning: could someone new to the account say it after one look at the top ad's headline? If so, the note describes the ad and not the buyer, so dig further. A learning that passes names something about the buyer, such as a fear, a boss they must answer to or a task they are tired of, and points to the words in the ads that show it.

The five rows below come from constructed data. They show the same test on five patterns.

Write each hypothesis on a card. This one works for any test:

HYPOTHESIS CARD
Date: [date]    Owner: [name]
Because [observation across ads: which ads, what they share, what the goal metric did],
we believe [cause: what this tells you about the buyer, written so a test can prove it wrong].
Next we test [concept] against [control], judged on [goal metric].
Evidence: [ad names or images] and one verbatim line: "[exact words from the ad]"
Basis: [platform], [attribution window or setting], [period], [results behind each number]
Held equal in both versions: [audience, budget, placement, start and end date]
Enough data means: [results needed per version, set before the test starts]
We drop this belief if: [the result that would prove it wrong]
If it holds: [what we do]. If it fails: [what we do].

For the invented campaign, the card reads: Because the two peer-number ads cost €53.49 per lead on 43 leads and the two feature-list ads cost €143.75 on 16 leads, and both peer-number ads open with a result said by a job title, we believe finance managers who sign for a whole team look for a peer with their own title before they weigh the result. Next we test P1 with the job title removed against P1 as it runs, judged on cost per lead. Both versions run at the same time with the same audience, budget and dates. Enough data means 40 leads per version. At 40 leads, chance alone moves each cost by about 16 percent, and the gap between the two versions by about 22 percent. We keep the belief only if the version without a title costs at least 50 percent more per lead than P1 as it runs, which is more than twice the chance swing of the gap. If it does not, we drop the belief.

What goes to the designer, the CFO, the client and the team?

Send the same facts in four cuts. Each cut answers the reader's private question in its first lines.

The finance cut uses three points from John Crabill at eLearning Industry: explain results in business terms and not in marketing terms, compare a percentage with a benchmark so the reader knows what it means, and lead with honesty when a goal was missed. The risk line and the date line are our additions. The designer cut is our reading. In the monthly creative review, performance and creative need to speak the same language, and pictures do that. So the designer's cut starts with pictures.

Here are the three cuts as templates.

Designer note: [campaign], [period]

Keep (repeat in the next batch)
[A thing visible in the frame, for example "the job title above the number"]. Seen in: [ad images]. What the buyer reacted to: [one sentence].

Stop
[A thing visible in the frame]. Seen in: [ad images]. What it did to the buyer: [one sentence].

Try
[What to make], against [control ad], because [cause].

Too early to say
[Angles or ads with too few results.]

I need from you: [format, size, date].

Subject: [campaign] creative result and the budget move I propose
To: [name]

Decision: move [amount or share] from [angle] to [angle] for [period].
Why: [angle A] costs [spend] / [results] = [cost per outcome]. [angle B] costs [spend] / [results] = [cost per outcome].
Basis: [platform], [attribution window or setting], [period], read on [date].
Results behind each number: [number] and [number].
If we are wrong, we lose: [amount, in money].
We will know by: [date], from [test].
What I need from you: [approval, or nothing].

Subject: What your ads taught us in [month], and what we do next

What we learned
1. [Learning in one sentence about the buyer.] [Picture of the ads.]
2. [Learning in one sentence about the buyer.] [Picture of the ads.]

What we do next
[The test in one or two sentences, and the budget shift.]

What we need from you
[Approval, asset or access, and the date.]

Open questions
[Only the ones that change a decision.]

Basis: [platform], [attribution window or setting], [period].

Send a PDF file to clients and finance readers: it opens anywhere and needs no login. Give the creative team a shared page that stays current. Keep the PDF and the log as your archive, so you can show what you knew and when.

How do you run a 30-minute creative review?

Send the page 24 hours before. Spend the 30 minutes on challenge and decisions, and not on reading the page aloud. Meet monthly. When a test ends, send the team a short note in between.

A common monthly creative review agenda has timed items of about 15 minutes of recap, 10 on top performers and 10 on under performers. Our agenda skips the recap because the page is the recap. It spends the time on groups, causes and tests. Idea generation belongs in a separate sprint meeting. In our reading, new ideas in the review slow the decisions.

CREATIVE REVIEW, 30 MINUTES
Monthly. The one-page note goes out 24 hours before.
Lead: [name]. Note-taker: [name]. Decision-maker: [name].

0:00 to 0:03   Basis and decision. The lead reads the decision line and the basis. Anyone who doubts the basis says so now.
0:03 to 0:10   Challenge the groups. For each angle, do all ads share one reason? Move ads between groups. Name the outliers.
0:10 to 0:18   Challenge the causes. For each angle, say the cause. Apply the junior test (could a junior say it after one headline?). Ask: what would prove this wrong?
0:18 to 0:25   Choose tests. Pick [two or three] hypothesis cards. For each one, fix the control, the goal metric, the enough-data rule, the owner and the start date.
0:25 to 0:30   Hand-offs. Designer: keep, stop, try. Decision-maker: the budget move. Note-taker reads the log rows aloud.

Rules
No new ideas in this meeting. Collect them for the idea sprint.
No ranking on any metric other than the goal metric.
Every test has an owner before the call ends.

Invite the media buyer or ads lead, the designer, the strategist or copywriter, and one person who can decide on budget.

How do you keep learnings from getting lost?

Keep a log with one row per test. Write the row on the day you decide. A learning that lives only in a deck gets lost. Read the log before every brief. Paste the rows for the angle you are briefing at the top of the brief. The team then does not test a dead angle again.

Date decided	Angle	Hypothesis	Test (concept against control)	Goal metric	Basis	Result (spend / results = cost)	Results behind it	Status	Decision	Ads
[date]	[angle]	Because [observation], we believe [cause]	[concept] against [control]	cost per [outcome]	[platform, window, period]	[spend] / [results] = [cost]	[number]	[open, supported, failed or too early]	[what we do]	[links to ad images]

Paste the block into a sheet. The columns split at the tabs. Four habits keep the log useful:

• Keep failed tests. A failed row removes a dead angle from the next brief.
• Use one status word for each row: open, supported, failed or too early.
• Write the date when a platform setting or a conversion window changed, because the numbers before and after do not compare.
• Search the log by angle before you write a brief.

Which numbers sit next to each learning?

Five items sit next to every learning: the goal metric with its two sums, the number of results behind it, the period with the read date, the attribution basis, and the platform. The goal metric is the cost of the outcome the campaign was bought for, such as a lead, an event registration or a video view. Compare an ad only with ads bought for the same goal. A reach campaign has no cost per lead.

Attribution basis

Every conversion count depends on a window. On LinkedIn you set a click window and a view window of 1, 7, 30 or 90 days on each conversion. LinkedIn's help page on conversion windows recommends 90 days click and 90 days view for most conversions and calls it the default in Campaign Manager. A reporting vendor's guide says LinkedIn typically counts 30 days for clicks and 7 days for views. The two sources disagree, so read the windows off each conversion rule in your own account.

Two more LinkedIn facts matter for reports. LinkedIn counts a member who converts several times in a window as one conversion, and it credits the latest click or view. Purchase and add to cart are the exception: LinkedIn counts each of those events. A changed window applies only to conversions counted after the change, so history keeps the old window. Write the date of any change in the log. LinkedIn made such a change itself in April 2025: event registrations now count 30 days after a click and 30 days after a view, up from 7 and 1. LinkedIn's help page says you may see more event registrations as a result.

On Meta, the attribution setting belongs to the ad set. Meta's results definition names 7-day click, 1-day view and 1-day engagement as the default. Meta's help on attribution models and settings describes standard, incremental and custom attribution. It says results in the Campaign Overview table cannot be compared across ad sets that use different attribution models. For a fair comparison, it points to the Compare Attribution Settings feature.

The same Meta page lists the standard settings. Click-through counts events within 1 day or 7 days of a link click. View-through counts events within 1 day of an impression. So two ad sets with a 1-day and a 7-day click setting count over different windows. Our reading: counts over different windows are not one fair number. Check the setting of every ad set before you add its ads into one angle. If the settings differ, print one sum per setting.

LinkedIn and Meta count on different windows. Give each platform its own column. Compare the order of the angles across platforms, and do not compare the sum.

Click and view on LinkedIn

In Campaign Manager, open the Ads tab, choose the Columns menu and pick Conversions & Leads. Three columns matter: Conversions, Click Conversions and View Conversions. LinkedIn's help on conversion reporting defines Click Conversions as actions completed after a click on your ad, and View Conversions as actions completed after the member saw your ad, even without a click. Conversions is the total. The same page says LinkedIn may use modeled conversions when it cannot tie a conversion to one member. If one angle's results are mostly view conversions, say so on the page. A reader who sees a cost per result may assume clicks. Then sort the angles on click conversions alone. If the order changes, the learning rests on view credit, and the page must say so.

The latest days

The newest days are incomplete. LinkedIn counts a conversion that happens inside the window you chose after a click or view (LinkedIn's help on conversion windows). Meta credits an action to an ad if it happens within a set number of days after a view or click (Meta's help on attributed actions). Our reading: a click or view can lead to a conversion up to one window later, so the counts for the newest days can still change. LinkedIn also says conversions you upload as a comma-separated values (CSV) file typically take 24 to 48 hours to reach reporting, and sometimes up to one week (LinkedIn's help on conversion reporting). Check once, with a conversion you know, which day your report books it on. Write the read date on the page. Mark the open days. Do not compare an open week with a closed one.

Few results

A count of results moves by about its square root from chance alone. This is general counting statistics, the Poisson distribution, where the variance equals the mean. It is a rough guide, and it is not a test. With 4 results the usual swing is 2, or 50 percent of the count. With 100 results it is 10, or 10 percent. In the invented table, the time-back group has 3 results, so chance alone moves its count by about 58 percent, and its CPL moves with it. The peer-number group has 43 results, so its count moves by about 15 percent.

Set your own enough-data rule before the period starts. Write it on the page. Mark every group below it as too early, and do not rank it. For single ads, when to pause an ad shows how to compare a gap with the noise.

How can Claude, ChatGPT or Nimblet build the first draft?

An assistant can group the ads, add the sums, quote the ad text and draft a first cause. A person must check the sums and sign the cause. Watch for four failures: wrong sums, an invented verbatim line, praise for an ad instead of a reading of the buyer, and missing data that the assistant hides. The first prompt closes each gap. Paste your export into it. The output format of the one-page note is inside the prompt.

You write creative learnings for a B2B paid social team. You write short, active sentences. You never invent a number, a quote or an ad.

I paste an export of ads below. Build the one-page creative learning note from it.

CONTEXT
Product and buyer: [one line: what we sell, to which role]
Platform: [LinkedIn or Meta]
Period: [start date to end date]. Read on: [date].
Campaign goal and the outcome it buys: [for example, a lead]
Attribution basis: [click and view windows, or the ad set's attribution setting]
Group column: [name of the column that holds the angle]
Minimum results for a group to count as read: [number I set before the period]

DATA (one row per ad: ad name, group, spend, results, opening line or ad text, and the attribution setting if the export has it)
[paste the export here]

RULES
1. Group the ads by the group column. Keep every ad in one group.
2. For each group, show total spend, total results and cost per result. Add spend and results first, then divide once. Never average the ads' own costs per result. An ad with spend and zero results has no cost per result: keep its spend in the group sum and also list it on its own line, "SPENT, NO RESULT".
3. Show the sums you used for every group so I can check them. Check that the group totals add up to the total of all rows I gave you. If you can run code, use code for the sums.
4. Mark a group "too early" when its results are below my minimum. Do not rank a too-early group.
5. Rank groups only against each other, on the one goal metric.
6. For each group, write one sentence on what the result says about the buyer that a newcomer could not guess from the top headline, and quote the ad line that shows it; if the ad text is missing, the group rests on one ad or the evidence is too thin, write "no cause yet" and name one extra cut of the data that would help.
7. End each group that has a cause with the Because line from the output format. A too-early group gets no Because line.
8. Start with a one-line decision. If the data cannot carry a decision, say what is missing.
9. State the attribution basis once under the title. Never combine numbers from different platforms or windows. If a row shows an attribution setting that differs from the basis I gave, keep that row out of the sums and list it.
10. No dashes. No filler words. Do not add facts that are not in my data.

OUTPUT FORMAT
CREATIVE LEARNINGS: [account or product], [period]
Read on: [date]. The latest days may still change.
Goal and outcome: [campaign goal]. Goal metric: cost per [outcome].
Basis: [platform], [attribution windows or setting].
Enough data means: [my minimum] results per group.
DECISION: [one line]
Then one block per group:
ANGLE: [name in the buyer's terms] ([ad names])
Result: [total spend] / [total results] = [cost per result]. Ads in the group: [number]. Too early: [yes or no].
What it tells us about the buyer: [one sentence]
Verbatim line: "[exact words from the ad]"
Because [observation across ads], we believe [cause]. Next we test [concept] against [control], judged on [goal metric].
Then one line each, in this order: OUTLIERS; SPENT, NO RESULT; TOO EARLY; OPEN QUESTIONS THAT CHANGE A DECISION.
Leave a field empty and say why when my data cannot fill it.

Suppose the export has one more column, the opening line of each ad. P1 opens with "[Result] in [time], says a [job title] at a [size] company." F1 opens with "What will your auditor ask for in week three of the close?" S1 opens with "Receipt capture, approvals, card feeds and [number] integrations." Suppose you set a minimum before the period. To pick one, take the smallest count at which the swing from chance, about the square root of the count, is a share you can accept, then check that most groups can reach it on your budget. On the invented table, the output looks like this. It is illustrative: the data, the product and the ad lines are invented, and bracketed parts stand for proof that a real ad would carry.

CREATIVE LEARNINGS: expense tool for finance teams of 50 to 200 people (invented), 1 to 30 September

Read on: 5 October. The latest days may still change.
Goal and outcome: the campaign buys a lead. Goal metric: cost per lead.
Basis: LinkedIn, conversion rule "Lead", click window [days], view window [days].
Enough data means: [your minimum] results per group. Groups below it are marked too early.

DECISION: Move October budget from the feature-list ads to the peer-number ads, and test the job title.

ANGLE: Peer number (P1, P2)
Result: €2,300 / 43 = €53.49. Ads in the group: 2. Too early: no.
What it tells us about the buyer: finance managers who sign for a whole team look for a peer with their own title before they weigh the result.
Verbatim line: "[Result] in [time], says a [job title] at a [size] company."
Because both peer-number ads cost less than half of each feature-list ad and both put a job title before the number, we believe the title carries the trust. Next we test P1 with the title removed against P1, judged on cost per lead.

ANGLE: Auditor's question (F1, F2)
Result: €1,700 / 22 = €77.27. Ads in the group: 2. Too early: no.
What it tells us about the buyer: the buyer may fear a surprise from the auditor more than a missing feature.
Verbatim line: "What will your auditor ask for in week three of the close?"
Because this group costs less than the feature-list group and more than the peer-number group, we believe the named moment helps and the peer's title helps more. Next we test F1 with a job title in the first line against F1, judged on cost per lead.

ANGLE: Feature list (S1, S2)
Result: €2,300 / 16 = €143.75. Ads in the group: 2. Too early: no.
Verbatim line: "Receipt capture, approvals, card feeds and [number] integrations."
No cause yet. Both ads list what the tool does and show no moment from the buyer's day. The extra cut that would help: compare these two ads with a feature ad that opens on a moment.

TOO EARLY: Time back (T1, T2), €450 / 3 = €150.00.
SPENT, NO RESULT: T2, €150.
OUTLIERS: none.
OPEN QUESTIONS THAT CHANGE A DECISION: do the peer-number leads fit the buyer we want?

Check every sum against your own arithmetic before you send. The second prompt tests the learnings an assistant or a teammate wrote. Paste it with a list of learnings.

Below are the learnings from our creative report. Test each one.

For each learning:
1. Write in one sentence what someone new to the account would say after one look at the top ad's headline.
2. If the learning says the same thing, mark it OBVIOUS. Rewrite it as a statement about the buyer, with only the evidence in the text I gave you. If that evidence cannot carry a deeper reading, say so and name the one extra cut of the data that would (for example, the same hook in two groups).
3. Mark it UNTESTABLE if no result could prove it wrong. Rewrite it as: Because [observation], we believe [cause]. Next we test [concept] against [control], judged on [goal metric].
4. Mark it OUTLIER if it rests on one ad.
5. A learning can carry more than one mark. If none applies, mark it OK and change nothing.
6. Do not add facts, numbers or quotes that are not in my text.

CONTEXT
Buyer: [role, and what the buyer risks by saying yes]
Goal metric: [cost per what]

LEARNINGS
[paste the learnings]

Answer as a table with four columns: the original learning, the junior version, the mark, the rewrite or the reason.

With Nimblet connected to your assistant over the Model Context Protocol (MCP), you can skip the paste. The assistant reads the account data itself, and the numbers and ad names come from Nimblet. MCP is on the paid plans. See the MCP page for how to connect.

Write the monthly creative learnings note for [account name], [period], for [reader: team, designer, finance reader or client].

Use my Nimblet account. First read how this account names its goals and conversions. Rank ads only against ads bought for the same goal, on that goal's metric. Group by [tag namespace, for example angle]. If the account has no such tags, group by the core message in the stored read of each ad. Use the stored read of each ad and the intelligence report to explain why a group works for the buyer, and quote a verbatim line from the ads as evidence.

Rules:
- Take every number and ad name from Nimblet, and show the ads with Nimblet's ad preview. Do not type a number you did not read in this chat.
- State the attribution basis under every conversion number.
- Mark a group "too early" when it has fewer than [minimum results per group, set before the period] results, and do not rank it. A too-early group gets no Because line.
- Start with a one-line decision.

Reader cuts. The team gets the full output format below. The designer gets keep, stop and try, each with the ad as a picture and one line on what the buyer reacted to, and no spend. The finance reader gets the decision, two costs per result with their sums, the basis, the risk in money and the date we will know. The client gets two or three learnings with a picture each, the next step and what we need from them.

OUTPUT FORMAT (full)
CREATIVE LEARNINGS: [account or product], [period]
Read on: [date]. The latest days may still change.
Goal and outcome: [campaign goal]. Goal metric: cost per [outcome].
Basis: [platform], [attribution windows or setting].
Enough data means: [my minimum] results per group.
DECISION: [one line]
Then one block per group:
ANGLE: [name in the buyer's terms] ([ad names])
Result: [total spend] / [total results] = [cost per result]. Ads in the group: [number]. Too early: [yes or no].
What it tells us about the buyer: [one sentence]
Verbatim line: "[exact words from the ad]"
Because [observation across ads], we believe [cause]. Next we test [concept] against [control], judged on [goal metric].
Then one line each, in this order: OUTLIERS; SPENT, NO RESULT; TOO EARLY; OPEN QUESTIONS THAT CHANGE A DECISION.

Nimblet also builds the report itself. It makes a document report: a branded PDF with the ads, the rankings and the reasons. You ask for it in the Nimblet chat with Nib, or from Claude or ChatGPT over MCP. ChatGPT needs a Business, Enterprise or Edu workspace with developer mode. The assistant writes the words and picks what to show. In tables, charts and ad lists, Nimblet fills every number, ad name and image from its own data when it prints. The AI does not type them there. The assistant may quote a number in its own text, so read those numbers yourself. You choose German or English and any period, and you can compare with the previous period or the previous year. A page that shows conversion numbers you mapped in Nimblet prints their attribution basis in its footer. The assistant is told to keep ads with too little data out of the verdicts and to mark them as too early.

You edit the text inside Nimblet before you send it. If you changed the text by hand, the assistant is told to keep your wording and tone when it edits near your words. Document reports are on the Pro, Enterprise and Agency plans. A report is made when someone asks for it, and Nimblet does not send one on a schedule. Nimblet only reads your ad accounts and never changes a campaign or a budget: you move the money in Campaign Manager or Ads Manager. The creative reports page shows an example, and pricing lists the plans. For the prompts that read single ads, see AI prompts to analyze B2B ads.