Read new-to-brand and path aggregates through their cohort rules, dataset coverage, and privacy limits before asking Claude for a story.
The blank row is not a shopper
AMC works with pseudonymized inputs and aggregate outputs under privacy safeguards. Narrow slices can be withheld or suppressed when they fail output eligibility rules. No returned row does not necessarily mean no eligible person followed the sequence. Do not use repeated micro-slices, external identity joins, or a Claude prompt to work backward toward individuals. Keep the result at the approved aggregate grain. A chart caption should say 'reported eligible paths under this query,' not 'every shopper journey.'
Before interpreting an absent group, compare the query definition and returned output: advertiser, dates, time zone, campaign grouping, filters, eligible conversions, and privacy constraints. Document whether a region or small campaign group fell out after a filter change. Claude can describe the rows actually present and note what is withheld or unknown; it cannot backfill a suppressed cell. Even a stable query only sees the signals the instance is entitled to use, not a complete off-Amazon history.
Define new-to-brand before comparing campaigns
'New to brand' (NTB) is not a universal count of new customers. In an AMC analysis it depends on available purchase signals, brand mapping, lookback, and the conversion definition. Record whether the question asks about NTB orders, attributed sales value, or an NTB proportion of eligible conversions. A brand identifier that omits part of a catalog can change the classification. The lookback and touch eligibility can also change the cohort. Put those rules alongside the result rather than hiding them in a query file.
Keep 1stPage’s Ads connection and AMC separate. Advertising_analytics can report an authorized new_to_brand advertising dataset; that is not an AMC path query or a row-level shopper history. For an AMC path discussion, provide a separately approved, privacy-eligible aggregate and its query definition. Ask Claude to compare the named sources without treating the 1stPage Ads result as the missing AMC table.
Imagine two permitted campaign aggregates: one has a high NTB fraction from a small order count; the other has a lower fraction but more NTB orders. A single percentage cannot tell the account team which campaign contributed more counted orders, much less which created incremental demand. Ask for each group's eligible numerator and denominator, plus any output limits. NTB classification describes the rule's observed cohort. It does not establish that a buyer would not have purchased without an ad.
Paths depend on where the clock starts
A sequence such as 'Sponsored Brands then Sponsored Products' exists only after an analyst defines eligible touch types, their ordering, the observation window, and the conversion event. Repeated touches need a rule. So do buyers with more than one eligible conversion. First-touch and last-touch summaries can both be internally consistent and tell different stories because they answer different allocation questions. Report the rule with the path distribution rather than offering a universal account of how people shop.
An advertiser may have sponsored-ads signals in its instance but no DSP signals. A chart without DSP exposure cannot show that DSP was absent from buyers' lives; it shows the dataset did not include that signal. Check the actual AMC instance, advertiser, regions, permissions, and dataset entitlements before asking for a cross-channel story. An Amazon seller login, Ads API access, and a Claude account do not by themselves grant AMC analysis. Never convert a product description into a claim that a specific account has access.
Two reports that disagree may both be right
Place the AMC aggregate beside the corresponding Amazon Ads report only after aligning advertiser, campaign set, currency, dates, time zone, event definition, and attribution window. Check data readiness and the AMC output's privacy eligibility too. If totals still differ, record each source's coverage instead of scaling one table until it matches the other. A reconciliation note should say which comparison is valid, which is only directional, and which cannot be made from the available signals.
Give Claude the approved aggregate and its query definition, not a raw identity-level export. Have it produce two columns: statements supported by returned aggregates and questions the output cannot settle. A path distribution can direct attention to a campaign group; an NTB split can identify a follow-up analysis. Neither reveals a named customer's route nor proves causal lift. That bounded analysis is more useful than a diagram that draws unseen or privacy-suppressed paths as facts.
Sources and boundaries
Check the current permissions and report definitions before acting on an answer. Example prompts are questions, not live account results.