Ask Google Ads account questions for consultants: choose the right measurement, scope and follow-up

Identify the question behind the wording

A consultant often receives a business question expressed as a short account question. Why did enquiries fall? Which campaign should get attention? Are the products we excluded still excluded? Each request needs a different kind of evidence. The most useful first step is to identify the decision the client wants to make, then define the relevant account selection.

Separate a request for a figure from a request for an explanation. How many primary conversions occurred in the selected period can be answered with a defined count. Why that count changed usually requires comparisons, campaign contributions and business context. A single total does not establish the cause of a movement.

Use the conversation's existing context. If the thread already names the account, campaign and period, a follow-up should preserve them unless the client changes them. Repeatedly asking which campaign after the campaign has just been identified makes the investigation slower and increases the chance of selecting a different target.

Clarify what is genuinely unresolved. A request for the worst campaign may need the business result definition. A request for a named campaign's enabled status does not need a broad discovery of the client's goals. Focused questions let the client supply the missing choice without restating the entire engagement.

Choose performance, settings or advice

Performance questions concern measured activity over dates: spend, clicks, primary conversions, selected value and their supported ratios. Settings questions concern current configuration: campaign status, a budget value, a targeting criterion or an asset relationship. These observations can complement one another, but they do not answer the same question.

An enabled setting is not proof that the resource has recently delivered impressions. A paused campaign may still have historical spend. A missing performance row does not establish that a resource was deleted. Choose the evidence corresponding to the client's request and state its date or configuration context.

Advice can use the established evidence and the client's business situation. Suggesting possible headline directions or explaining a previously retrieved result does not inherently require another provider call. A factual request about current account state does. The framework should distinguish creative reasoning from an account read rather than routing every conversation into a data query.

Keep the conclusion within the evidence. A reported destination topic is not a live website crawl. A product targeting state is not Merchant eligibility. A conversion count is not a count of qualified customers unless the selected measurement and business evidence support that meaning. The consultant should make the boundary explicit at the point it matters.

Set the account and parent scope

Establish the connected account before retrieving account-specific evidence. A channel's connection determines which account the AI can access through its available tools. If the required source is not connected, the response should politely explain the missing connection rather than pretending that an empty result represents the account.

Name the campaign when the question is campaign-specific. For an ad group or keyword, retain the actual parent campaign and relevant group. Names provide understandable context; internal resource references keep the exact selection stable behind the scenes. A client-facing answer should not require raw account or asset identifiers to be useful.

Resolve duplicate names deliberately. Two groups named Seasonal Offers in different campaigns are different targets. If the conversation cannot distinguish them, show the relevant parent names and ask for the intended selection. Guessing from the most recently displayed example is unsafe when the request concerns a live action.

Scope also includes status and channel choices. A request for enabled campaigns differs from a request for all campaigns that spent during a period. A product question may concern PMax asset groups, Standard Shopping ad groups or both. Preserve the selected population instead of applying an invisible convenience filter.

Resolve dates and observation basis

A performance request needs actual dates. Resolve last thirty days as the intended rolling range, last month as the calendar month and an explicitly named interval as that interval. Include the start and end in the explanation. The client should not have to guess whether the system substituted one convenient period for another.

Use the source's relevant time zone when establishing day boundaries. A request for today is a request for incomplete current activity; a request for completed days excludes that unfinished part. Neither should be labelled as the other's population. If a comparison crosses sources, preserve the relevant timing context before discussing any mismatch.

The conversational conversion measurements default to conversion-date reporting unless the client explicitly asks for ad-interaction dates. That choice affects which results fall inside the interval. Do not answer a conversion-date question with an interaction-date total and then explain the difference as an unexplained tracking error.

Product relationship reporting has its own historical-performance note. Current targeting is joined to the requested historical observation, and the listing performance uses its documented interaction-date basis. The answer should preserve that distinction. A current exclusion does not establish that the product was excluded throughout the entire historical interval.

Use the shared field and measurement catalogue

The framework discovers registered fields, business measurements and related datasets before preparing a report. This provides a reusable path for different questions rather than a separate handwritten tool for every phrasing. A request for a campaign count and a request for a filtered product membership use declared fields and relationships appropriate to their selected dataset.

Local settings discovery describes readable fields and entity relationships without performing a provider read. Actual settings retrieval then selects the live evidence required by the question. Discovering a capability and observing an account are separate operations; a catalogue entry does not prove that a particular account contains a matching resource.

Google's query-language overview distinguishes resources, attributes, segments, measurements and field metadata. That distinction supports a deliberate evidence request: choose what belongs together, then retrieve the selection needed for the answer.

The consultant does not need to write raw query syntax in Slack. Normal questions can use the reviewed business measurements and relationship datasets. A raw provider field request remains a separate explicit choice. It should not silently replace a business definition whose primary-action rules or selected-event requirements differ.

Select the business result definition

Total conversational Ads conversions count primary actions by default. Secondary actions are excluded unless explicitly requested. Total primary value can include valued actions outside the purchase category. These are intentional definitions, not an instruction to repair the advertiser's measurement or remove actions that the client deliberately made primary.

Purchases select primary purchase-category actions. Leads combine the primary lead categories without separating lead stages in the business output. Calls select primary call actions and can have value. Offline conversion measurements expose uploaded or imported actions; they can overlap the purchase, lead or call groups and should not be added to them as a separate extra pool.

A named action selection answers a more specific question. The client can request the count or value of Contact Enquiry rather than all leads. Inspect available action metadata when needed and preserve the chosen name. An explicit request for a secondary named action can include it without changing the normal default for other reports.

For ratios, select compatible numerator and denominator definitions. Purchase value divided by cost is different from all primary value divided by cost. Total Analytics purchase revenue divided by selected advertising cost is blended ROAS or MER, with the spend source stated. The framework should not hide that choice behind a generic return label.

Select the corresponding Analytics event

A site's lead event can have a business-specific name. When comparing Ads leads with Analytics, obtain the event the user names or confirms. The AI can retrieve a bounded list of observed event names and suggest the most plausible one, but a plausible name is not proof that the event records the intended outcome.

The distinction matters for calls. A phone-link click can show intent to contact a business; it does not establish that a completed call happened. If the client chooses that event for comparison, retain the two series' meanings. The chart should not relabel the click event as a verified call completion.

Named Ads conversion comparisons similarly require the selected Analytics event. A request that explicitly names Contact Enquiry and enquiry_submit supplies the intended mapping. If the client asks to see likely lead events first, show available names and ask for confirmation before treating a suggested candidate as the comparison target.

Current release controls restrict named Analytics comparison paths whose delivery verification is not complete. Explain that specific availability boundary if it applies. It is different from an unavailable connection, absent event data or an invalid field combination. Precise wording prevents an application capability limit from becoming a false claim about Analytics.

Define best and worst before ranking

The worst campaign depends on the outcome the client values. For a store, a supported value-to-cost measure may be appropriate. For a lead programme, cost per selected lead may be more relevant. A campaign with high spend is not automatically the worst, and a campaign with one inexpensive result is not automatically the best.

Specify the measurement, selected population, period and any minimum evidence requirements. Ranking an account's enabled campaigns differs from ranking every campaign with historical activity. If the question needs a purchase definition, use that definition rather than generic primary value simply because it is available.

Preserve missing and undefined results. A campaign with no usable denominator cannot be ranked as an ordinary zero ratio. A campaign with unavailable data cannot be described as the lowest performer. The framework's measurement rules and coverage should guide how the result is presented and where clarification is required.

After retrieving the ranking, explain the leading result and the relevant qualification. A request to pause it then becomes a new action proposal against the exact resolved campaign. The ranking supplies evidence for the decision; it does not constitute live-edit permission or prove that pausing will improve the account.

Apply exact conditions and logical groups

A consultant can ask for values above a boundary, at least that boundary, below it or at most it. These comparisons are distinct. More than three excludes a value of three, while at least three includes it. Use the condition corresponding to the client's wording instead of treating all threshold requests as a loose greater-than filter.

Supported conditions also cover equality, inequality, selected lists, ranges and appropriate text matching. Between uses inclusive numeric bounds. Text conditions can distinguish contains, prefix and suffix checks. Validate the field type and operator before retrieval; an invalid filter should not appear to succeed merely because the result happens to be empty.

Logical groups express different selections. Products with return above three and spend above a chosen boundary must satisfy both. Products matching one category or another satisfy the selected alternatives. Negating a condition should preserve unknown values as unknown; missing performance is not evidence that the opposite threshold is true.

Operator support remains dataset-specific. Joined relationship results reject regular-expression filters and offer supported exact or text alternatives. A universal-looking syntax catalogue does not mean every operator can run on every provider path. Explain the supported replacement instead of silently dropping part of the client's selection.

Preserve unknown values

A missing value carries information about the evidence, not just the layout. A product without matching historical performance cannot be confidently assigned a return ratio. A campaign whose source query failed cannot be treated as having no spend. Keep unavailable, undefined and measured zero states distinct.

Joined conditions preserve unknown through logical composition. If return is unknown, a request for return above three cannot include the row as a verified match. Negating that test does not turn the unknown value into proof of return at or below three. The answer should state how many observations could not be evaluated when that coverage affects the result.

Calendar gaps require the same care. A missing additive point may be filled with zero only where the report is complete, the provider suppresses zero rows and the framework's timing conditions permit it. A failed read cannot supply a series of zeros. Undefined ratios cannot be repaired by filling their numerator or denominator arbitrarily.

Present the available evidence and the remaining task. The client may narrow the scope, choose a more suitable period or obtain the missing connection. A concise limitation is more helpful than a numerical answer that hides the failed prerequisite. It also prevents a subsequent edit from being based on an invented selection.

Count the complete selected population

Use the complete-count path for how many questions. Displayed examples are often intentionally capped for readability, so counting the preview is not a valid account total. The framework can count the selected dataset and provide a breakdown using declared grouping fields when that is the requested operation.

Name the unit. A count of campaigns differs from a count of keyword rows, listing branches or product targeting occurrences. The same product in two enabled groups can contribute two occurrences. A unique product inventory count answers a different question and should not replace the occurrence count without the client's request.

Do not aggregate repeated-membership performance as though every row represented independent revenue. Relationship reporting preserves a declared identity and rejects unsafe duplication. A grouped membership count can be valid while summing the attached performance across those same memberships would be misleading.

A work limit is not a partial answer to a complete count. If the eligible selection exceeds the framework's safe read or row budget, narrow the campaign or group scope and explain the boundary. Never claim completeness after silently dropping the excess rows or treating the displayed portion as the account population.

Keep hierarchy and relationship identity

A relationship question needs the actual parent chain. A keyword belongs to its ad group and campaign; a PMax listing belongs to its asset group and campaign. An ad group is not the parent of another ad group simply because a generic formatter has reused the wrong field. The registered dataset defines the relationships used for the answer.

The framework keeps resource identity internally and presents meaningful names and context. This lets it resolve a selected target consistently without displaying raw identifiers to the client. When a product identifier is the subject of the question, it can be relevant product data; that is distinct from exposing internal account or resource references in ordinary prose.

Related rows are joined on the declared keys. A relationship that should return one matching row cannot safely return several and multiply the original facts. The executor rejects that fanout rather than manufacturing inflated counts. A missing required relationship key likewise needs an error or a narrower investigation, not a guessed parent.

Use path context where it helps the client understand an exclusion or setting. A product under a category and brand branch has a more informative explanation than a disconnected list of IDs. Preserve the path's actual structure and group ownership instead of reconstructing a plausible hierarchy from names.

Ask about product targeting occurrences

A product can be included in one enabled group and excluded in another. For an exclusion-occurrence question, preserve those separate memberships. PMax uses asset groups; Standard Shopping uses the relevant ad groups. Each group can have its own listing tree, even when several groups contain the same inventory products.

The product membership dataset handles explicit item IDs and products covered by broader branches. A category or brand exclusion can affect products without an individually named leaf for every item. Matching current inventory to the relevant tree supports an occurrence view, while the listing-node dataset answers a different structural question about branches and partitions.

By default, the membership population uses enabled groups within enabled campaigns. A specifically selected campaign-status scope can change the campaign selection while the enabled-group requirement remains. State the actual scope in the answer. Do not include paused groups when the client asked for the enabled-group exclusion population.

Report a count with its campaign and group context: the selected excluded occurrences across the relevant PMax campaigns and asset groups, plus Standard Shopping campaigns and ad groups. The framework's summary can distinguish those populations. A combined total should still name the unit rather than implying that it counts unique inventory products.

For an individual campaign or group request, apply that scope before drawing the conclusion. A result for one asset group cannot establish the campaign-wide exclusion count when other enabled asset groups have independent trees. The consultant should retain the requested parent selection throughout the investigation.

Combine current targeting with historical performance

A request for excluded products with historical return above three combines two observations. Current listing membership supplies the targeting state. The selected historical performance supplies the return inputs. The answer should describe both rather than imply that the product was excluded for the whole measured interval.

The relationship reader obtains performance only when selected fields or conditions require it. A simple count of current exclusions does not need a historical performance call. Asking for return, spend or another declared performance field introduces that dependency. This is a framework decision based on the selected data, not an extra scan performed for every product question.

Product performance is joined using the relevant declared identity. If matching history is unavailable, the ratio remains unknown. A product with zero spend does not have a normal return-on-spend value. Such rows should not be quietly assigned zero and included in the opposite side of a threshold test.

The same product can count in several selected groups while its measurement relationship needs careful handling. Do not sum repeated membership observations into an invented account revenue total. The framework distinguishes membership counts from performance aggregation and reports the requested occurrence population.

Use the result to ask the next practical question. A historically productive product may currently be excluded because it is unavailable, no longer profitable or outside the offer plan. The account evidence can identify the selected occurrences; the client's business context determines whether a supported inclusion change should be proposed.

Follow an invented investigation

Consider an invented store, Cedar Hollow Supplies. Its consultant examines one PMax campaign with two enabled asset groups. Both groups contain the same three products. The first group excludes the second product; the second group excludes the second and third products. The selected result is three excluded product occurrences, representing two distinct product IDs.

The client asks how many exclusions belong to the first group. That follow-up changes the group scope, so the answer is one occurrence in the invented example. It should not repeat the campaign-wide count or remove the group context. If another campaign is selected later, that is a new population to retrieve.

The client then asks which excluded occurrences had return above three in a named historical period. Suppose two have usable historical ratios of four and five, while one has unavailable matching history. The strict greater-than-three condition yields two known matches. The unknown observation is identified as unevaluated, not called a low-return product.

If the client asks to include those two, the framework prepares the exact complete selection for review. The consultant does not copy a short preview into an edit and assume it represents the full result. The proposal names the relevant products and group parents, while the current-state checks protect the live operation from stale targeting evidence.

This example illustrates population and condition handling, not customer performance or a recommendation to re-enable every historically productive product. The business must still confirm availability and suitability. The account question establishes a selection; it does not replace the client's commercial decision.

Preserve context through follow-ups

A follow-up should modify the established plan deliberately. Show that as a line chart changes presentation. Use the last seven days changes dates. Only the enabled groups in this campaign changes scope. Compare it with a selected Analytics event adds a measurement mapping. Preserve the other resolved choices unless the client changes them.

Reference the actual selected target when the client says it or that campaign. An earlier executed change can provide context for a subsequent request, but a new operation still needs its own prepared proposal. Remembering the target does not grant continuing permission for every future edit.

When a follow-up conflicts with an earlier definition, explain the effect. Changing from all primary value to purchase-only value alters the measurement, not just the label. Switching from all Analytics traffic to Ads-attributed traffic alters the population. The next answer should state the new selection instead of implying continuity that no longer exists.

Ask for clarification only where the thread cannot resolve the intended change. If several campaigns have been discussed, a pronoun can be ambiguous. Offer the relevant names and parent context. If the last selected campaign is clear, asking which campaign again adds friction without improving the evidence.

Data references have a scope and lifetime within the request workflow. If an earlier dataset is no longer available, reconstruct the validated plan and retrieve the required evidence. Do not fabricate a remembered numerical series or claim that a lost data reference proves the account itself is unavailable.

Request the appropriate native visual

Ask for a visual when it helps answer the question. The AI can render requested provider data as a native Slack chart or table using the returned data reference. A numerical answer does not automatically need a chart, and an unrequested visual should not add provider work or clutter the conversation.

A combined chart preserves separately defined series. Compatible money measures can share a requested graph with clear source labels; a count and a money value need a presentation that preserves their units. The current native output permits up to three selected series, so a larger request needs a focused selection.

For a longer period without explicit daily detail, the framework can select a coarser readable grain. State whether the points are weekly or monthly. If the client explicitly requests daily observations, preserve daily detail and explain a necessary split. A single chart preference does not justify discarding an explicit breakdown.

Native delivery has twenty points per series per panel and a maximum of two chart panels. Longer explicit detail can require a shorter interval or a supported table. The answer should explain that output boundary plainly. Do not substitute an image chart when the intended output is native Slack.

Use readable labels and no more than two decimals. Current-year dates can omit the year, while older dates retain the context needed to distinguish them. Weekly points should describe actual date intervals. The explanation must still retain the report's complete selected period.

Keep creative conversation conversational

A consultant may ask for five keyword phrases around an offer, possible headline directions or a clearer explanation for a client. These requests can be answered creatively from the supplied context. They do not automatically ask for measured search volume, account performance or a live account change.

Distinguish a suggested phrase from a researched measurement. If the client asks for traffic history, competition or bid estimates, the available keyword research tools can provide the relevant evidence with the required connected account. If the client asks only for wording ideas, explain them as suggestions rather than attaching unrequested numerical claims.

Creative suggestions also have no implied upload. A proposed responsive search ad can be discussed and revised in conversation. Creating it in the account requires a supported action, the complete required inputs and owner confirmation. The client should see the actual target and creation status before approving.

A request to explain an existing report can reuse the established result. The AI does not need another performance query merely to make the wording more suitable for a nontechnical reader. If the client asks whether the account has changed since that observation, the new factual question does require relevant current evidence.

Read failures accurately

A structured planning error is not a provider limitation. An unknown field, incompatible measurement combination or unsupported filter should lead to the appropriate selection or clarification. Repeating the identical non-retryable call cannot make that plan valid and should not become the response strategy.

Connection failures, permission failures and empty successful results are also different. A missing source connection needs the connection workflow. A denied account read needs access review. A complete query with no eligible rows needs an explanation of its scope. Each requires different wording and a different next step.

Safe work limits preserve completeness. If the selected account population exceeds a reader's supported budget, narrow the campaign or group scope. Do not turn a truncated result into a confident total. A client can still obtain useful evidence through a smaller defined selection without being told that the provider has no such data.

State the failed prerequisite specifically. The phrase could not confidently read that answer is not enough when the thread contains a clear parent campaign. Determine whether target resolution, stored context, field validation or the provider call failed. The user-facing reply should then identify the choice or action that resolves the actual problem.

Avoid unnecessary account work

Begin with local catalogue discovery where the question concerns available fields or settings. It describes capabilities without reading the account. Retrieve the actual account values only after identifying the required selection. This reduces speculative provider requests while retaining the evidence needed for a factual answer.

Reuse a returned dataset for its text explanation and requested visual. The same totals should not be re-read separately for a sentence, table and chart. Compatible requests can share the retained evidence, while changes of period, population or required current state justify new reads.

Relationship dependencies are selected by the requested fields and conditions. A current membership count does not automatically retrieve product performance. A requested return filter does. The executor keeps the relevant keys internally and validates that the enrichment does not duplicate the base facts.

Filter at the appropriate stage. Conditions on base fields can scope the membership selection; conditions on related performance belong after the required relationship data is available. Unsupported query pushdown must not drop an OR branch or change the logical request. The exact framework condition remains the authority for the selected result.

Refresh when freshness is part of the question. Reuse is appropriate for explaining an existing observation, while a request about a newly changed setting needs a current check. The consultant should remove duplicate work, not substitute stale evidence for a live-state question.

Move from a finding to an approved edit

A read can identify a candidate for action, but execution uses a separate prepared proposal. The AI selects a supported action and resolves its exact target. For a deeper asset, the proposal includes the relevant parents. Each logical step is shown plainly so the client can understand the intended account change.

For dataset-derived selections, use the framework's complete edit selection rather than copying the displayed preview. If the selection exceeds the review boundary, narrow it before preparing the proposal. Silently editing only the first few displayed products would contradict a request concerning all matching occurrences.

The connected owner confirms before the live API call. The pending proposal states that no changes have been made yet and expires after fifteen minutes. Current relevant settings are checked again before execution. The AI's successful retrieval of a count is not continuing authority to mutate every counted resource.

Dependency ordering is planned locally where the supported action requires it. A creation chain needs its parent resources before its children. That ordering does not prove universal atomicity across every service or make every read-only field writable. Use the actual supported edit capability for the selected setting.

Use audit findings and wasted-spend evidence as supporting investigations where relevant. They help define a decision, while the live action still needs its own target, proposal and confirmation.

Keep execution and later outcomes separate

After confirmation, the progress message should visibly indicate that the change is in progress and remove the pending control. The completed message then reports the actual execution outcome. A visible update is useful, but the status must come from the workflow rather than a predicted success.

Executed and verified means the supported result was checked in the account. An operation sent without complete verification has a different status. An uncertain dispatched write requires reconciliation and must not be blindly replayed. The consultant should explain the remaining task instead of asking the client to repeat the same request casually.

A verified setting does not prove performance improvement, provider approval or website functionality. Those outcomes need their own evidence. Enabling a selected group establishes its configuration; it does not promise traffic. Changing a URL establishes the checked stored field; it does not certify the customer journey.

Return to performance when the client asks about effect. Use the relevant later dates and the same agreed measurement definition, with any changed population explained. Keep the original action record separate from the later commercial observation so the client can assess both accurately.

Leave a usable consulting record

Record the resolved question, source, dates, population, metric and principal result. A successor should understand what the answer measured without reverse-engineering internal tool arguments. Campaign and parent names make the scope readable, while exact identity remains available to the workflow behind the scenes.

Preserve the unresolved question and owner. An unavailable relationship read, unconfirmed Analytics event or uncertain execution needs a specific next step. A confident narrative should not erase those dependencies. If the client chooses to defer an action, record that business decision without presenting it as a verified account improvement.

Connect the question to client reporting when the evidence becomes a recurring business review, and to monitoring when a configured observation should recur. A one-time answer does not automatically create a saved report or monitoring rule.

The reusable framework provides fields, measurements, relationships, conditions and output paths that can evolve together. Its value is the ability to answer new combinations without inventing a dedicated tool for each sentence. The consultant's responsibility remains to select the correct meaning and keep the conclusion within the returned evidence.

Answer common account-question concerns

Does every conversation require an API call?

No. Creative suggestions and explanations can use supplied context. Current account facts need relevant provider evidence.

Can the AI count a preview?

A preview can show examples, but an account total needs the complete-count path and sufficient coverage.

Does at least include the boundary?

Yes. Greater-than excludes it, while greater-than-or-equal includes it. Preserve that distinction in the filter.

Does the same excluded product count once across groups?

An occurrence count preserves each selected enabled-group membership. A unique inventory count is a different request.

Can the AI choose a likely lead event?

It can suggest one from observed names, but the user must name or confirm the corresponding event before it is treated as the selected mapping.

Does a read authorise a live change?

No. A supported edit needs an exact prepared proposal and connected-owner confirmation.