Is an LLM Too Slow for Attribution? Check the Data Lag First

Usually not. The model is one of four clocks between an order and a decision, and the one you can set: Artificial Analysis, a third-party benchmarking company, lists total response times for one model, Claude Sonnet 5.5, from 6.81 seconds at its medium setting to more than 400 seconds at its max setting (leaderboard read on 30 September 2026). The other clocks run in hours and days: Google says GA4 data processing can take 24 to 48 hours, and credit for a key event can change for up to 12 days.

What do measurements say about each clock?

1. The model clock. Artificial Analysis lists total response times for Claude Sonnet 5.5 at its medium, high, xhigh and max settings, each labelled "with fallback" on its page: 6.81 seconds, 18.40, 37.40 and more than 400. The leaderboard is live and its figures change, so treat these as dated. Its methodology page defines end-to-end response time to include input processing, model reasoning and answer generation. This is the benchmarking company's own measurement, published by that company and not independently audited, and it is not your prompt. OpenAI's latency guide (official documentation) says generating tokens is almost always the highest latency step.
2. The data clock. Google says the daily BigQuery export typically lands mid-afternoon in the property's time zone but can be delayed to the next day. Streaming export is best-effort and excludes traffic-source data for new users and new sessions (official documentation).
3. The credit clock. Google says attribution credit for a key event can change for up to 12 days after the event.
4. The test clock. Lewis and Rao (Quarterly Journal of Economics, 2015, peer-reviewed) studied 25 field experiments with major US retailers and brokerages. The median confidence interval on return on investment was over 100 percentage points wide, and they conclude that informative advertising experiments can easily require more than 10 million person-weeks (people multiplied by weeks of exposure).

Do rate limits slow a workflow?

Anthropic measures its limits in requests, input tokens and output tokens per minute for each model class, and OpenAI also counts requests per day and images, at organisation and project level (official documentation). Exceeding an Anthropic limit returns a 429 error. Compare the numbers on your own limits page with the calls your workflow makes per minute.

How do you measure your own lag?

1. Data lag. Note the timestamp of a recent Shopify order. Record when it first shows in GA4's Realtime report and when it appears in your BigQuery daily table (the table details show "Last modified").
2. Credit lag. Save yesterday's revenue by channel now, again a week later and again after two weeks. Note which days moved.
3. Model lag. Time one real question from send to a number you can act on, including your reconcile step.
4. Decision cycle. Write down how often you move budget.

Pass: the slowest clock fits inside your decision cycle. Fail: the slowest clock is longer than your cycle, so a faster model buys nothing.

OpenAI's guide lists "Don't default to an LLM" as a principle, and it applies here: once a query is written and checked, the saved query runs without a model call. An LLM is a good way to write it.

Later, once the lag is clear, a causal attribution read like Causality Engine's takes a GA4 export and returns its results in minutes. It is as current as the export you give it, so the 24 to 48 hours above still apply.

Sources, 30 September 2026: LLM Leaderboard (Artificial Analysis, read 30 September 2026; published by the benchmarking company, not independently audited); Benchmarking Methodology (Artificial Analysis, 2026); Latency optimization (OpenAI, 2026); Data freshness (Google, 2026); BigQuery Export (Google, 2026); The Unfavorable Economics of Measuring the Returns to Advertising (Lewis and Rao, Quarterly Journal of Economics, 2015); Rate limits (Anthropic, 2026); Rate limits (OpenAI, 2026).

Related answers

• Can an LLM Read the Time Series in Your Attribution Data?
• Why an LLM Gives Different Answers to the Same Data Question
• Holdout test calculator
• Geo testing, explained
• Causal attribution, explained