An 11,000-row search-term report is not a better brief for Claude. On a large ad account, it can be more data than the assistant can reason over well, wrapped in an answer that looks complete.

Say I’m reviewing a Google Ads account spending $20k a month. I ask Claude to pull every search term from the last 30 days and flag the waste. Seconds later: a clean table and a confident total. Neither tells me whether it actually analyzed the whole report.

The fix starts before the answer: filter, aggregate and rank inside the tool call, then drill in. Asking for everything and sorting it out in chat puts the hard work in the wrong place.

The context window is the working surface, not the account

I think of the context window as the desk Claude gets to work on. Questions, loaded tool definitions and returned ad data all take up space. On a small account, that may be fine. On a big one, the desk can fill before the useful analysis starts.

Here is the running example: a Google Ads account with 40 campaigns, 800 ad groups and 11,000 search terms over the last 30 days. Through groas as my MCP connector, I ask Claude to “pull all search terms and find waste.” Depending on the fields returned, that result can run to hundreds of thousands of tokens.

The report is not the only thing competing for room. One published example describes three MCP servers consuming 143k of a 200k-token window before the first question. A connection to the account does not give Claude unlimited working space.

The desk analogy has a limit: fitting the data is not the same as analyzing it correctly. A report can be incomplete, and a complete report can still be calculated badly. I need to check both.

A polished answer can hide an incomplete report

An overloaded answer often looks finished. The warning signs are in what it leaves out:

  • It lists 50 search terms when the account contains thousands, without labeling them as a subset.
  • It names four campaigns but misses the biggest spender.
  • It gives a total that I cannot reconcile with the account.

The last one is especially dangerous. A tidy dollar figure feels more authoritative than a partial list, even when both come from the same incomplete view.

In one published audit, an assistant reported $1,840 in waste, or 61% of budget, after counting the same cost twice for terms appearing in two ad groups. Checking the rows brought the figure to $640, or 21%.

That example is an accounting failure, not proof of a context-limit failure. The distinction matters. Narrowing the result helps Claude inspect it, but I still need the grouping and calculation to be right.

A confident total is a claim to verify, not evidence that every row was read. Before changing a bid, I want to know what went into it.

Why the full report may never reach Claude

The size of the account and the size of the tool response are different problems. The MCP specification itself sets no size limit on tool results, and its built-in pagination covers listing tools, not the data returned by an individual tool call. A tool therefore needs its own way to bound or page a large report.

If it returns too much, the application around the model may intervene. Claude Code, for example, has been documented as warning past 10,000 tokens of MCP output and capping it at 25,000 by default, saving longer output to a file rather than putting it all in chat.

That protects the window. It does not mean the assistant has analyzed the saved file.

Reported tests describe oversized results being truncated, failing transport, or reaching the model as a partial list. In one test involving 24,000 invoices, a cap passed only 10 rows to the model with no way to fetch the rest. One model identified the sample; another multiplied it up to estimate a total.

This is where a plausible answer becomes misleading: the assistant treats the visible slice as the whole account.

My first defense is a bounded result, not a more forceful request to “analyze everything.” The next five patterns put that into practice.

Five prompting patterns that shrink the work before it reaches chat

Each pattern changes the report Claude receives, not just the wording of its answer. That distinction is the point. Asking Claude to display 20 rows after retrieving 11,000 does not solve the same problem as returning 20 rows from the tool.

I specify the reduction in the tool call, then check that the returned result reflects it. A prompt is an instruction, not proof that the tool performed the operation.

1. Filter and group before returning rows

For the $20k account, I do not start with every search term. I start with a narrower question: which terms spent more than $25 without a conversion in the last 30 days?

I also specify the grouping. A search term can appear in more than one ad group. If I want a term-level review, I need spend and conversions summed by term before applying the threshold.

My prompt:

For the last 30 days, group search terms by term and sum spend, clicks and conversions. Return only terms with summed spend over $25 and zero summed conversions. Apply the grouping and filters in the tool call.

This follows the principle of filtering, projecting and aggregating before listing. The irrelevant rows stay out of the working context rather than arriving with instructions to ignore them.

Define the date range, grouping and threshold before asking for the list. Otherwise, “find waste” leaves too much of the report’s shape unspecified.

2. Build blended ratios from summed totals

Filtering controls volume. Aggregation also controls the math.

Suppose one campaign spends $100 to make $800 in revenue: 8.0x ROAS. Another spends $900 to make $1,800: 2.0x. Averaging those ratios gives 5.0x. Dividing total revenue by total spend gives $2,600 / $1,000 = 2.6x.

The second calculation answers the account-level question. The first gives both campaigns equal weight despite their different spend.

For the running account, my prompt is:

Return total spend and total revenue for the last 30 days, computed from the full matching dataset. Compute account ROAS as total revenue divided by total spend. Show both totals and the formula.

That is the useful part of defining the metric and formula before computing it. I am not handing Claude 40 ratios and hoping it chooses the right way to combine them.

Compute blended ratios from summed totals, never by averaging the component ratios. A smaller answer is useful; a correctly defined answer is essential.

A tiny desk buried under search-term reports, with one highlighted sheet pulled out

3. Rank the result and cap it explicitly

A filtered list can still be too long. Next, I give it an order and a limit.

For an initial waste review, I want the highest-spend candidates first, not whichever rows happen to arrive first:

For the last 30 days, group search terms by term and sum spend and conversions. Keep terms with spend over $25 and zero conversions. Sort by spend descending and return the top 20 only. Apply the limit in the tool call and state whether more matching terms exist.

That turns an open-ended request into a bounded review. It also makes the omission explicit: this is the top 20, not every matching term and not total account waste.

Good pagination makes the boundary visible through a limit, a cursor and an indication that more results exist. When those are available, I want Claude to report them rather than quietly presenting page one as the whole result.

A ranked subset is useful when it admits it is a subset. I use it to choose where to investigate, not to declare the audit finished.

4. Drill down through the account structure

The running account has 40 campaigns, 800 ad groups and 11,000 search terms. I do not need all three levels at once to decide where to look first.

I work down the structure:

  1. Return campaign-level totals for the last 30 days: spend, conversions and CPA.
  2. Select the two campaigns with the worst CPA for closer review.
  3. Return ad-group totals inside those campaigns only.
  4. Inspect search terms inside the three worst ad groups.

Each answer narrows the next question. Instead of carrying the entire account through the conversation, I carry the decision about where to look.

Paging through everything is not automatically a practical alternative. At 10 rows per page, 24,000 rows require 2,400 calls. A technically pageable report can still be the wrong workflow.

This route also has a boundary: inspecting selected campaigns does not establish what is happening everywhere else. I label the scope of the finding as carefully as the finding itself. “These campaigns need review” is not “the rest of the account is clean.”

5. Keep arithmetic in the tool or code

The second pattern defines the right formula. This one specifies where the calculation happens.

Even a correctly defined total is unreliable if Claude reconstructs it from a long list in chat. I want the tool to return the summed values, or the data processed in code, rather than asking the model to add rows in prose.

In a reported 12.4MB test, Claude, ChatGPT and Cursor got the right checksums only when they saved the result to a file and queried it in code instead of reading all the rows in chat.

For the campaign view, my prompt is:

Return one row per campaign for the last 30 days, with summed spend and summed conversions. Compute CPA as spend divided by conversions. Use the full matching dataset for the sums, not just the displayed rows.

That final sentence separates an account calculation from a page calculation. Twenty displayed rows may be enough for inspection; they are not automatically enough for an account total.

I want quoted totals to trace to a tool result or a reproducible calculation. A polished sentence is not a calculator receipt.

The three-line check before I act

Before acting on a big-account answer, I ask Claude to provide three things:

  1. Coverage: the date range, matching row count and any limit or truncation.
  2. Calculation: the exact formula and where it was computed.
  3. Evidence: the aggregate rows behind the total, plus the detail needed to inspect a flagged item.

I do not ask it to paste the entire report back into chat. That would recreate the problem. I want enough information to verify the scope and calculation, then drill into the relevant rows.

When I use groas for marketing MCP workflows, I keep that order in the prompt: establish coverage, define the metric, compute from totals. If the answer cannot show those basics, I check Google Ads before touching a bid.

The same sequence applies outside search terms. For 5,000 products, I would first make sure ChatGPT is reading real ad data rather than guessing. Then I would request summed spend and revenue by product for the last 30 days, minimum $50 spend, sorted by loss descending, top 25 only. After that, I would drill into the top three.

Connecting the data is step one. Keeping the question narrow is step two. Neither replaces the other.

Apply the model to a six-location CPA spike

Now take a different case: a local services advertiser with six locations wants a full Q3 review across Google and Meta, every campaign, every keyword, plus Reddit mentions to explain a CPA spike.

That request fails the desk test before the first result arrives. It mixes account totals, detailed performance rows and unstructured text. I would split it into four passes:

  1. Locate the change. Return summed spend, leads and cost per lead by advertising channel for Q3, one line per channel.
  2. Narrow the campaign set. Return the top 10 campaigns by spend increase versus Q2, with the increase computed in the tool.
  3. Inspect search detail. Within the two worst Google campaigns identified in the review, return search terms spending over $50 with zero booked jobs. Keep the result bounded and state whether more matches exist.
  4. Examine customer language separately. Return Reddit quotes about price and scheduling from the last 90 days, capped at 20.

The first pass establishes the totals. The next two narrow the performance investigation. The last adds customer language without burying the account data under a pile of quotes. Those quotes may help interpret the spike; they do not establish its cause on their own.

Each answer now has a job, a scope and a boundary I can inspect. I can follow a number back to its aggregate instead of trusting a summary of an oversized export.

Small accounts grow. Date ranges widen. The habit of asking for everything survives until the month it breaks.

My rule is simple: if my prompt has no date range, no threshold and no limit, I rewrite it before sending it. Filter in the call, sum before listing, rank with a number and drill in steps. Give Claude a question it can answer from the data it actually receives, not a report it has to pretend it read.