How to Review Social Post Performance With Claude

Analyze social media post performance with Claude by supplying a defined set of posts, checking which metrics are available, and comparing similar posts on the same measure. Use the result to choose one content change to test, while keeping missing data and uncertainty visible.

A useful review ends with a decision you can explain: “I will try another worked example because this one received more recorded saves than comparable posts.” It should not end with an unsupported verdict that a topic guarantees growth or that engagement proves a sale.

Follow these six steps for a review you can repeat alongside your weekly content planning.

What can Claude conclude from your available post data?

Start by giving the assistant actual evidence. Anthropic documents CSV uploads and code execution for data analysis. You can use an export or an authorized connection, but first ask Claude to identify what it received rather than assuming it has inspected your account.

Separate three kinds of output:

• Recorded facts: the supplied post has 12 saves and a stated observation timestamp.
• Calculated comparisons: its save count is twice the median of the selected comparison set.
• Interpretations: its concrete checklist might be worth trying again.

The numbers in those examples are illustrative. Only the first two statements could be verified from a suitable table alone. The third needs the actual post and remains a hypothesis about why the result happened.

We make Distinctful, a social media scheduler for entrepreneurs with or without a newsletter. Its MCP tool reference includes get_post_performance, which returns available metrics, unavailable fields, observation timestamps when recorded, and comparisons with usual performance when eligible. Those records help ground a review; engagement does not establish revenue, and the product does not send newsletters.

If you use a file, remove private details the task does not need. Keep the post identifier or public link so you can trace each conclusion back to its evidence. Supply full post text separately if the export contains only a shortened title.

Try this opening prompt:

Review this dataset before analyzing performance. Identify the account, publication period, available metrics, metric definitions, missing values, and observation timestamps. State whether this is a complete export or a selected slice. Do not infer sales, diagnose an algorithm, or change any posts.

How do you choose one platform, period, and comparable post set?

Write the question before choosing the winning post. For example: “Which of my recent educational text posts received more recorded comments?” That is a narrower and more answerable question than “What content works?”

For a first review, choose one account on one platform. Define the publication period, format, and whether paid distribution is excluded. Then list the fields you need:

Field	What to record
Post identifier	A stable ID or public URL
Content	Full text or a reference to the complete post
Publication time	Date, time, and timezone
Format and purpose	Text, image, or video; education, invitation, or another job
Primary metric	The exact provider field you want to compare
Observation time	When the supplied metric was recorded, if known
Distribution notes	Paid promotion, collaboration, or another known difference

Treat this as a preparation checklist, not a promise that every export supplies every column. If a field is absent, name the limitation and narrow the question.

Ask for the number of included and excluded posts and the reason for each exclusion. A list of top posts cannot tell you how typical posts performed unless it also contains a suitable baseline. Likewise, a filtered result cannot establish what happened outside its filter.

Keep posts with different jobs visible but separate. An invitation to book a consultation and a quick educational tip need not be judged by the same reader action. Decide which question each group can answer before combining them into a ranking.

How should you handle missing metrics and different observation times?

Preserve the difference between a recorded zero and an unavailable value. Zero saves means the supplied field records zero. An empty save field means you do not have that measurement. Replacing the empty field with zero would create a result the source did not provide.

Ask Claude to return a short data-quality note before the ranking:

Count the available and missing values for the chosen metric. Preserve zeros. Identify duplicate post IDs, missing timestamps, and posts outside the selected period. Explain which rows can be compared without filling unknown values with estimates.

Also distinguish publication time from observation time. A post published on Monday and measured on Tuesday has a different measurement window from one published on Monday and measured on Friday. For cumulative counts, comparing those snapshots as if both describe the same elapsed period would answer a different question.

Choose a consistent elapsed-time window when you have the necessary snapshots. If you only have current totals, describe the comparison as totals observed at the available times. Do not ask Claude to reconstruct a post's first-day performance from its later total.

When observation time is missing, freshness is unknown. A recent export date does not, by itself, establish when every underlying metric was last observed.

If a particular post is absent from a ranked response, request that post directly. Absence from a shortlist is not a diagnosis of missing analytics or a failed connection. Our guide to why posts reach few followers covers the separate distribution question; keep the data check and the explanation of reach distinct.

Which posts beat your usual performance on the same measure?

Define “usual” before calculating a multiple. A practical descriptive baseline is the median of an explicitly selected set of comparable earlier posts. The median is the middle ordered value, or the mean of the two middle values for an even-sized set; NIST explains why extreme observations affect it less than the mean.

For a positive baseline, calculate:

Performance multiple = post's metric value / baseline median for that same metric.

Report the metric, comparison-set size, inclusion rules, and observation window beside the result. If the baseline is zero, do not report a division-based multiple. Show the absolute values instead. If the comparison set is empty, there is no baseline to calculate.

When a tool supplies its own comparison, preserve its metric and rules. Do not rename a reach comparison as an engagement comparison or describe a general account baseline as a format-matched experiment. A supplied multiple and your manually filtered comparison can legitimately answer different questions.

Next, inspect the full content of the higher-performing posts. Ask Claude to point to the passage supporting each proposed pattern. A shortened title does not establish what the post taught, what the video showed, or what readers reacted to.

The guide to finding outlier posts and developing original ideas offers a useful next step: extract the reader question and useful approach without copying someone else's expression or treating popularity as proof of causation.

What would a clearly labeled illustrative review conclude?

Consider a fictional consultant comparing recorded saves on educational posts from one account. The five earlier comparison posts have save counts of 2, 4, 5, 7, and 12, all captured after the same elapsed period. Their median is 5.

These are invented teaching values, not customer results, benchmarks, or measurements from a live account.

Candidate post	Recorded saves	Measurement condition	Defensible conclusion
A: kickoff checklist	10	Same elapsed period as baseline	2 times the baseline median on saves
B: process opinion	0	Same elapsed period as baseline	Zero recorded saves in this snapshot
C: project example	Unavailable	Observation time unknown	Cannot rank it on saves
D: approval template	15	Longer elapsed period	Exclude from the equal-window comparison

A careful review would say:

Post A exceeded the selected earlier-post median on recorded saves. Its full text contains a checklist, which is a possible feature to test again. This small comparison does not establish that checklists caused the difference or that any post produced a customer.

It would not call Post D the winner of the equal-window comparison, treat Post C as zero, or turn the result into a universal posting rule. Even Post A's apparently promising result remains a reason to investigate.

Check the calculation yourself: 10 divided by 5 is 2. Then check the more consequential judgment: whether the rows belong in the comparison at all. Correct arithmetic cannot repair an unsuitable comparison set.

How do you turn one observation into a testable content change?

Choose one change small enough to evaluate. For the fictional consultant, that might mean adding a worked checklist to the next educational post while keeping its audience, topic family, and intended action similar.

Write a compact test note:

1. Observation: the kickoff checklist exceeded the selected median on saves.
2. Hypothesis: a concrete checklist may make this advice more useful to revisit.
3. Change: include a short checklist in another relevant educational post.
4. Measure: recorded saves at the same elapsed observation window.
5. Review: compare with the stated baseline and record other differences.

This is a practical content experiment, not a randomized trial. Keep alternative explanations visible and avoid declaring a lasting rule from one follow-up result. If the comparison remains weak, the useful decision may be to collect consistent snapshots before changing your plan.

Ask Claude for a proposal rather than an automatic publishing action:

Suggest one content change supported by this review. Cite the post evidence, state the uncertainty, and define the metric and observation window. Draft a brief for my review; do not schedule or publish anything.

Carry the approved brief into your weekly social content plan. Save the original question, selected data, calculation, and decision beside it. At the next review, you can assess what happened against the expectation you actually wrote, instead of inventing a new explanation after seeing the numbers.