.svg)
Both models got every calculation right, but their first budget advice differed and their final advice matched. Claude first suggested moving money from TikTok to Meta; ChatGPT suggested keeping a 50/50 split and waiting. After two follow-up questions, both models recommended not cutting TikTok.
Below is how to build a Meta and TikTok report in a single chat and which prompts to use. Then comes an experiment: how far you can trust the conclusions of such a report. The situation is a familiar one: two platforms, the same budget, and a manager who sees that a Meta customer costs a third less. We connected our Meta and TikTok ad accounts via MCP and built this report for one month. Then we gave the table to Claude and ChatGPT, asked each model the same three questions in a row, and checked every answer: the calculations, the statistics and the logic behind the recommendation.
The table shows our own campaign results, and the AI answers are real too. We did not edit them, only shortened the retelling. Steps 1–3 describe Claude's answers; ChatGPT's answers to the same questions follow in a separate section.
How to build a Meta and TikTok report in one chat
There are three ways to bring data from both platforms into one AI chat, from a one-off export to a permanent account connection. Connections run over MCP, a protocol that gives an AI assistant direct access to an ad account.
The connectors are still in beta, and the setup terms and list of supported chats keep changing, so check the platform's help centre before you start. Connectors can not only read reports but also create and edit campaigns. Start with read-only access.
Next, here is how to build the report itself, step by step.
1. Start with one platform
Connect your Meta ad account and ask for a summary:
Show the results of the Meta ad account for [period]: spend, impressions, CPM, link clicks, CTR, CPC, conversions and cost per conversion. One table, metrics in rows.
Check spend and the number of conversions against Ads Manager. If the numbers differ, specify the period, time zone and conversion event in the prompt. Once everything matches, your single-platform report is ready. You can stop here and compare two campaign groups or two periods.
2. Add the second platform
Now pull the same metrics for TikTok for the same period and add them as a column to this table. Use link clicks on both platforms, one currency and one time zone. If the platforms calculate a metric differently, note it below the table.
The last sentence of the prompt matters. In Meta it is easy to take "Clicks (all)" instead of link clicks, and then CTR and site conversion stop being comparable. That is exactly what happened in our table — more on that in the experiment.
3. Add leads and customers from your CRM
Ad accounts see conversions but do not know which leads became customers. Upload a CRM export of leads with UTM tags for the same period, or connect the CRM if it has a connector:
Here are the CRM leads with UTM tags for the same period. For each platform, count leads, customers, lead-to-customer conversion and CAC. Add these rows to the table. Show leads without UTM tags as a separate row.
Count leads from the CRM, not from the ad accounts: both platforms can claim the same lead.
4. Save the prompts as a template
The result is a table like the one in "Experiment: the data" below. Ours was built exactly this way: platform metrics through connectors, leads and customers from the CRM. Save the three prompts in your project instructions or notes, and change only the period each month.
Before you compare the platforms, check two more things:
- both platforms use the same conversion definition;
- campaign names show which campaigns are retargeting and which target cold audiences.
Prompts for analysing the report
Once the table is ready, move on to analysis. These prompts work with both exports and connected accounts. For a single platform, replace "Meta and TikTok" with "two campaign groups" or "two periods".
The "Checking a conclusion" and "Context" prompts are easy to skip once the summary is ready. The experiment below shows what happens if you stop at the first answer.
Experiment: the data
One month, the same budget on both platforms, campaigns optimised for leads. These are our own Meta and TikTok campaigns: platform metrics were pulled through connectors, leads and customers come from the CRM.
Each stage of the funnel has a different leader. TikTok wins on reach and clicks, the platforms are almost even on leads, and Meta leads on cost per customer. This is the table a manager receives along with the question of where to move the budget.
Step 1. A plain question
We uploaded the table to a new Claude Opus 5.5 session with no history or instructions and asked the question a manager would ask:
Here are one month of ad results in Meta and TikTok. Compare the platforms' performance. Where should we reallocate the budget?
Claude did not rush to name a winner. It split the funnel: TikTok wins on reach and clicks, Meta on conversion to customers. It then pointed out on its own that a single customer accounts for the CAC gap: one more TikTok customer and the cost per customer would be equal. It also found a bottleneck: 11,250 clicks produced only 29 leads. Claude suggested that the Meta report might show all clicks rather than link clicks. We checked, and it did.
We checked every claim in the answer against the source table.
The arithmetic is flawless. The weak spot is elsewhere: Claude called the data insufficient and still advised moving 20% of the budget. A reader who scrolls down to "What to do" will only see "60/40".
Step 2. How confident are you?
The second prompt was short:
How confident are you in this conclusion? Is there enough data to make a budget decision?
Claude revised its own answer. It withdrew the 60/40 recommendation and stated plainly that keeping 50/50 is just as justified. It also called its earlier threshold of "30–50 leads and 8–10 customers" insufficient: such numbers will become more stable but won't prove the difference.
We recalculated the statistics ourselves.
Claude's conclusion held up: you can't decide on the basis of the CAC gap. The errors were in the details — sample size and the width of TikTok's interval. They don't affect the decision, but they shouldn't go into a report for a manager without recalculation.
One more finding: Claude suggested comparing the platforms not with each other but with the acceptable cost per customer. That question changes the task itself.
Step 3. Adding context
At the third step, Claude received what a marketer usually knows but an export doesn't show. The click and CRM details here are real: the Meta report shows all clicks, there were 1,500 link clicks, and leads and customers were counted in the CRM. We modelled the rest — the attribution model, the retargeting share, the sales cycle, the average order value and the margin — to see how the models would react to a typical situation.
- In the Meta report, the "Clicks" column is "Clicks (all)". There were 1,500 link clicks. The TikTok figures are link clicks.
- Leads and customers were counted in the CRM by UTM, last-click model.
- 40% of the Meta budget went to retargeting site visitors, including people who first came from TikTok. TikTok ran on cold audiences only.
- A deal closes in 3–4 weeks on average. Of this month's leads, 5 from Meta and 6 from TikTok are still open.
- Average order value is $600, margin 50%, almost no repeat purchases.
What changes in your conclusions? Where should the budget go next month?
The recommendation turned around completely. Claude now advises not to reallocate anything: both platforms pay off, and the last-click model hands part of TikTok's contribution to Meta retargeting. Cut TikTok, and you also shrink the audience that Meta retargets. Claude itself called its 60/40 advice a mistake.
What changed after the recalculation:
- Traffic. With link clicks, a Meta click costs $0.33, not $0.10. TikTok is four times cheaper, and its link CTR is three times higher. At Step 1, Claude called TikTok's CTR "slightly lower" and has now corrected this itself.
- Site conversion. Meta converts at 1.0% versus 0.22% for TikTok. The difference is significant, but it compares warm visitors with cold ones, not platforms. Claude dropped its claim of low site conversion on both platforms: the problem is only with cold traffic from TikTok.
- Open deals. Among closed deals, conversion is 30% for Meta and 25% for TikTok. If the remaining deals close at the same rate, CAC will be about $110 and $140.
- Payback. A customer brings $300 in margin, so the break-even CPA is $75–90 against the current $33–36. Even without the open deals, the overall CAC is $200.
- A new idea. Split Meta retargeting into visitors from TikTok (by utm_source) and everyone else. That shows how many Meta leads TikTok actually prepared.
Checking the calculations:
There is one point we would soften. Claude writes that both platforms "pay off with room to spare" and that there is no need to wait for a winner. But TikTok's closed-deal CAC is $250 against a $300 threshold — a thin margin. With five customers, the 95% interval for the overall cost per customer runs from $86 to $616. Whether the result will repeat isn't proven yet. The advice to add budget gradually, by 20–30%, and to stop if CPA exceeds $50 takes this uncertainty into account.
We covered how to grow a TikTok budget while keeping cost per result under control in our guide TikTok Smart+ Cost Cap.
The same questions in ChatGPT
We sent the same three prompts to ChatGPT — the GPT-6 Astra model in Medium mode — in a new chat. ChatGPT first went through Step 3 with an assumed figure of 1,900 link clicks. It received the real 1,500 in a separate message and recalculated the traffic. There were no arithmetic errors here either. The models made mistakes in different places, and after receiving context they arrived at the same budget plan.
Checking ChatGPT's calculations:
After the correction, ChatGPT pinpointed the key fact: spend, leads and sales didn't change, so CPA and CAC stayed the same and there's no new reason to move budget. It kept its $500 / $300 / $200 recommendation.
ChatGPT is more cautious in its conclusions; Claude pays more attention to the data. ChatGPT gave no hasty advice but missed "Clicks (all)" — a real quirk in our data that later turned the traffic comparison upside down. Claude spotted it right away but twice sounded more confident than the data allowed: in the 60/40 advice and in its estimate of TikTok's contribution.
What held up
Both models' arithmetic held up entirely. Claude's budget recommendation did not: over three steps it went from "move to Meta" to "don't touch TikTok".
The experiment produced four useful observations.
- AI calculates better than it decides. Not a single arithmetic error in any answer. The inaccuracies were in sample size and interval estimates, and the recommendations turned out to be the weak link.
- "How confident are you?" works. With no new data, Claude withdrew its 60/40 advice on its own, and ChatGPT clearly separated the month's facts from the forecast. Ask this question after every budget conclusion.
- Context matters more than the export. What reversed the decision wasn't a new calculation but what the ad accounts don't show: click type, retargeting, the sales cycle, margin. Correcting 1,900 to 1,500 clicks changes neither CPA nor CAC — ChatGPT rightly pointed this out. Claude answered more decisively with the new figure, but that may simply be run-to-run variation. Without this context, AI analyses a table, not a business.
- The models make mistakes in different places. One is more cautious, the other more attentive to the data. If a decision costs money, ask a second model the same question and compare the answers.
Limitations: one run each in Claude Opus 5.5 and GPT-6 Astra (Medium) on one month of data. Asking again may produce a different answer. We re-ran Step 3 for Claude once we learned the real number of link clicks; ChatGPT received it in a separate message.
Before you send the report to your manager
What to add to the context prompt:
- Which clicks the report shows: all clicks or link clicks. Column names in Meta and TikTok don't mean the same thing.
- Where leads and customers come from: the ad accounts or the CRM. Which attribution model is used.
- What share of the budget goes to retargeting and whom it reaches.
- How long a deal takes and how many of the period's leads are still open.
- Average order value, margin and repeat purchases — the acceptable cost per customer is calculated from these.
And three checks before you send it:
An AI chat saves time on the summary and the calculations. The budget decision stays with the marketer who knows what's behind the numbers.
.jpg)