How to audit an AEO data export

Before you sign a monitoring contract, ask for a small export from the package you intend to buy. A screenshot can demonstrate a dashboard; a usable export lets another person check the claim. This guide proposes an editorial acceptance checklist. It does not report a completed vendor test or a new panel result.

01 / Define the unit before counting

Write down what one row represents: a configured question, a captured answer, a citation, or an aggregate. Keep separate files when those units differ. One answer may cite several URLs; a prompt summary may combine multiple engines and dates. Counting all of those rows as answers inflates the denominator.

Writesonic documents separate prompt and answer views and a table-export facility. That documentation alone does not guarantee a complete answer-and-citation export on every plan. Ask for the file and inspect the fields rather than inferring them from the interface. Vendor field descriptions .

02 / Request a joined sample

Our suggested request: export 12 fixed prompts for one agreed market and date window, together with every available answer and citation. Ask for stable prompt ID, exact prompt text, run ID, capture timestamp and timezone, engine or surface, model where known, country, language, collection method, answer text, cited URL and run status. Mark unavailable fields as unknown. These are acceptance criteria, not a claim that every platform supplies them.

Use a composite key such as prompt ID + engine + market + capture timestamp + run ID. Check that every citation maps to an answer and every answer maps to a prompt. Preserve the original files and save transformations separately. If a product provides only daily aggregates, label them as aggregates and narrow the analysis you promise.

03 / Reconcile the run ledger

Hypothetical example, not test results: 12 prompts × 2 engines × 7 daily captures means 168 planned answers when there is one capture per combination. If the ledger contains 160 successes, five failures and three missing records, account for all 168. Retrying a failure creates a new attempt; it should not silently double the number of scheduled observations. Do not turn missing records into successful answers with no mention.

Market duplication can also change the plan's prompt allowance. Writesonic counts identical questions separately across country-language markets. Record quota consumption separately from the number of captured answers. Market accounting documentation .

04 / Keep the dataset and cadence visible

A discovery index and a fixed custom-prompt panel answer different questions. Ahrefs documents search-backed index data separately from custom prompts, whose freshness depends on the chosen schedule. Tag which dataset the export came from before comparing dates or products. Ahrefs' dataset and freshness explanation .

Keep engine, market and branded status available as filters. Compare the same cohort across periods. If you change the prompt library, keep a stable subset for trend reporting and describe the added questions separately. A change in coverage can alter an aggregate even when no individual answer changes.

05 / Recompute one metric

Define a mention as a brand appearing in answer text and a citation as a linked source under your stated URL rules. Review redirects, subdomains and third-party pages explicitly. Record whether an answer with no sources is valid, failed or out of scope. Give a teammate the written definitions and ask them to calculate one metric from the exported rows.

Hypothetical calculation: 24 cited answers out of 160 successful in-scope answers is a 15% answer citation rate. Using 168 scheduled observations produces approximately 14.3%, a different denominator that includes failures and missing records. Either can be reported if accurately labeled; do not compare them as if they were the same metric. Never use the number of citation rows as the number of cited answers.

06 / Decide whether the evidence is portable

Accept the export for your intended analysis only if another person can join the records, explain missingness and reproduce the selected metric. Ask about retention, pagination, row limits, API charges and whether exports remain accessible when the contract ends. A restricted export can still support a narrow workflow, but its limits belong in the buying decision.

Keep the run configuration, raw files, audit notes, definitions, calculation and change log together. This makes a later review possible without reconstructing history from a chart. Vendor-reported improvements and a before-and-after observation should remain separate from a claim that a tool caused the change.

Use this checklist with our Writesonic research review, then connect the evidence to a measurement plan. The site publishes new scored panels only when the dated run and reviewer records are complete. Writesonic GEO review . Citation measurement guide . Buying guide . Panel evidence requirements .

Sources checked 5 October 2026. No commercial relationship with Writesonic or Ahrefs was disclosed to this reviewer. The worked examples and acceptance criteria are our proposed workflow, not vendor performance findings. Commercial disclosure . All field guides .