Ask Google Ads account questions for agencies: get a scoped answer in the client conversation

Define the agency question-handling job

An agency receives account questions from clients, specialists and colleagues preparing reviews. The practical task is connecting each question to the right account, evidence definition and next decision. Fast answers are useful only when they retain that meaning.

A question might concern measured results, current settings, existing assets or general advice. These requests need different responses. Reading an account metric does not answer a creative brainstorming question, and a general explanation does not establish the client's current configuration.

Consider a fictional agency, Rowan Harbour Media, managing several client accounts with similarly named campaigns. A colleague asks how many campaigns are enabled, while another asks what caused a cost increase. The first needs a current scoped count; the second needs defined performance evidence and careful interpretation.

A consistent process starts with intent and scope rather than a fixed list of API calls. The assistant should select the supported information needed for the question and explain any missing requirement. It should not retrieve unrelated datasets merely because several tools are available.

The agency benefits when a reply can be checked and continued. A useful answer states what was read, which population it concerns and what follows from it. It keeps an unresolved interpretation open instead of making the account sound more certain than its evidence allows.

Keep the connected account explicit

The conversation needs the appropriate connected account context. A campaign name alone is not enough across an agency's clients. Similar labels can describe unrelated assets, and a result from another channel should not become evidence for the current account.

Confirm the account name where it helps the user assess the answer. The client-facing reply should use recognisable context rather than expose account identifiers unnecessarily. Internal resolution and human-readable naming serve different purposes.

If the connection is missing or no longer reachable, explain the access requirement politely. The assistant should not invent a result from a remembered account or treat an access failure as proof that there are no campaigns or conversions.

Keep supplied business context attached to the right client. A statement about the service area, product offer or budget agreement can guide interpretation, but it is not a provider fact verified simply by being mentioned in the channel.

At staff handover, retain which account the question concerned and what remains to be checked. This small record is especially useful when the agency repeats the same question across several clients. It prevents a valid answer from being reused in the wrong context.

Identify what the user is asking for

Classify the request by its practical purpose. A measured question asks for account evidence. A settings question asks for current configuration. A creative request asks for ideas or wording. A change request asks for a supported operation prepared for human review.

The same subject can appear in each category. Suggest five relevant keywords is creative discussion. Give traffic estimates for these keywords needs research evidence. Add these keywords to a named group is a live-change request with target and confirmation requirements.

Do not treat tool availability as a reason to use it. The assistant can explain general concepts and draft copy from supplied facts without retrieving provider data. Conversely, it should not answer account-specific numbers from memory or a plausible estimate.

Identify combined requests explicitly. A user can ask for ideas and measured volume, or for a performance read followed by a requested proposal. Each part needs its own evidence or authority boundary rather than an undifferentiated response.

This intent distinction keeps the conversation natural. The agency can discuss possibilities, verify account facts and prepare actions in one thread while the assistant remains clear about which statements are creative suggestions, retrieved evidence and unexecuted proposals.

Clarify consequential ambiguity

Ask a short question when an ambiguity would change the answer or operation. Which conversion outcome, which similarly named campaign or which corresponding Analytics event can be consequential. The assistant should not guess merely to avoid a clarification.

Do not ask for facts already established in the usable context. If the conversation resolved one campaign and the follow-up clearly refers to it, preserve that target. Repeating a generic campaign-selection question can discard useful context and frustrate the user.

A performance question may need a selected measure. The framework has an automatic performance interpretation, but an explicit purchase, lead or named-action instruction should override an unrelated default. Keep the user's intended definition visible.

For commercial facts, ask the person who can supply them. Whether a client sells a product or serves an area may not be discoverable from the account. Another provider read cannot substitute for a missing business constraint.

Clarification should move the task forward. State the missing choice and its effect plainly. A useful question lets the user resolve the plan, while a long technical questionnaire often adds effort without improving the evidence.

Discover supported evidence

The conversational reporting framework exposes field discovery, settings discovery, data retrieval and requested rendering. Field discovery can describe provider compatibility, business measurements and registered datasets. It gives the assistant a reusable way to select evidence rather than a dedicated tool for each phrase.

A catalogue describes what can be requested; it is not proof that the connected account has observations for every entry. Actual account facts need a successful read. An available field with no returned data should not become an invented result.

Settings discovery is a local catalogue step and does not itself read the provider. The assistant can use it to understand entities and relationships before requesting the relevant current values. This avoids treating every dependency question as another API call.

Some discovery requests do read account metadata, such as conversion actions or observed Analytics event names. That evidence helps resolve a named selection, but an observed name does not independently establish its full implementation or business meaning.

Choose the smallest supported plan that answers the request. A question about a campaign's current status does not require all historical keyword metrics. A question about a daily trend needs actual compatible daily observations rather than an unrelated period total.

Choose the business measure

Use the reviewed business measurements for ordinary conversion and revenue questions. Aggregate primary conversions includes all selected primary actions. Primary purchases and primary leads use their corresponding groups. Named conversions resolve a particular action by name.

Primary conversion value can include value from non-purchase actions. Purchase conversion value is narrower. The assistant should preserve that difference when the user asks about revenue or ROAS instead of relabelling every conversion value as purchase revenue.

Call and offline questions also have supported conversion context. Primary call actions can carry value, while uploaded and imported actions can appear in the offline selection. These groups can overlap other categories and should not be added as independent populations.

Secondary actions are excluded by default unless the user requests them. Configured primary aggregation does not deduplicate separate actions that record overlapping business activity. The report follows the selected account definition rather than repairing tracking design implicitly.

Explain the choice near the result. A short label such as primary purchases or the named action helps the user assess what the answer means. Technical precision should make the conversation clearer, not bury it in raw provider field names.

Define the population and conditions

State which entities or traffic population the question concerns. Account, campaign, group, keyword and product relationships are different scopes. A broad result should not be substituted for a specific target because it is easier to retrieve.

Preserve the requested status selection. Enabled entities describe a current configured population; paused entities can still have historical performance. Parent status can also matter, so an entity's own enabled label should not imply that its full path is enabled.

Use the user's comparison boundaries exactly. Higher than and higher than or equal to differ at the threshold. Conditions joined by both differ from conditions joined by either. Negation also changes the selected population.

Keep the conditions in the answer where they aid interpretation. A count of keywords with a minimum selected cost and zero selected purchases is a defined selection, not a universal wasted-spend total. The commercial conclusion needs additional context.

When a condition requires unavailable performance evidence, retain that limitation. An unknown value should not become zero so the row qualifies. A generic framework is useful because it preserves these semantics across different supported datasets.

Preserve period and reporting basis

Use actual dates for performance questions. Last month and the last thirty days are distinct periods. A current budget cycle can have another boundary. The assistant should retain the selected period in follow-ups unless the user changes it.

Keep the source's time-zone context for complete-day reporting. A colleague's local yesterday can differ from the account or property boundary. Date labels should make the selected observation window assessable.

Daily conversion outcomes default to conversion date in the current conversational framework. An explicitly requested interaction-date view has a different meaning. Do not switch between them silently when explaining a trend.

Google's data-freshness guidance describes different processing timelines and later updates. Recent measured results can therefore need a qualification without being treated as a tool failure or a confirmed business decline.

A visual-grain change should preserve the dates. Weekly grouping can make a long trend readable; it does not authorise replacing the request with a different week range. Explicit daily detail should remain explicit in the supported plan.

Answer counts from complete selections

A count question should use a complete supported selection. A preview of ten campaigns or keywords is for readability, not proof that only ten exist. The framework's count semantics should establish the selected population rather than count displayed examples.

Name the unit. Campaigns, asset findings, product occurrences and unique inventory items are different counts. A client-facing number should not change its meaning simply because the assistant shortens the explanation.

For product listing questions, the same product can occur in several enabled groups with different inclusion or exclusion decisions. A group-specific occurrence count is not automatically a unique product total. Broad grouping nodes are also different from individual product identifiers.

Apply scope and conditions before interpreting the count. A campaign-specific answer should not reuse an account total without a compatible selection. A performance-filtered count needs the requested metric and comparison boundary.

If coverage is capped or unavailable, explain it rather than claim completeness. The user can narrow the selection or choose another supported report. A truthful bounded answer is preferable to a confident total derived from a short preview.

Read parent and relationship context

Many account questions concern relationships rather than a single scalar value. A keyword belongs to an ad group and campaign. A Performance Max listing path belongs to an asset group, while Standard Shopping uses an ad-group context.

The registered dataset framework can describe supported hierarchical and relationship reporting. It gives the assistant a way to select the proper parent path instead of guessing it from names or implementing a special tool for every wording.

Repeated names need resolution. A group called Main can exist under several campaigns. For a requested operation, the human should see recognisable target and parent names that establish which entity is intended.

Do not invent missing parents. An account-wide asset finding may not include every attachment context. The assistant should obtain the supported information needed for the question or state the limitation plainly.

Relationships also determine action order. Creation of a child requires the appropriate parent context, while a change to an existing child may need no new parent creation. The dependency framework should prepare the required sequence without unnecessary account-wide reads.

Handle source comparisons clearly

A comparison can use related Ads and Analytics measures with distinct source labels. Different attribution does not automatically make every comparison invalid. The assistant should check the supported measurement families and selected populations.

Purchase-related value, purchase counts, all primary conversions and confirmed selected-event comparisons have their own definitions. A lead or call comparison needs the corresponding Analytics event named or confirmed by the user.

The assistant can list observed event names and suggest a plausible correspondence, but the suggestion is not confirmation. A phone-link click is not automatically a completed call, and a form-like event name does not prove the client's intended lead implementation.

Keep compatible compared series separate. They can appear in a readable visual without being summed into a combined result. Overlapping activity across sources is not proof of additional revenue or unique conversions.

If the user wants all-traffic Analytics revenue divided by an agreed cost source, identify the view as blended ROAS or MER. Resolve the spend source where needed instead of silently assuming the same attribution as an Ads return measure.

Separate unavailable evidence from zero

A returned zero is an observation; an unavailable field or failed read is a gap. The assistant should not tell the user the account has no activity simply because it could not obtain the requested evidence.

Structured plan errors also need accurate handling. A failed field combination is not proof of a universal provider limitation. Use the error to identify the needed selection and avoid repeating an unchanged non-retryable call.

Keep partial coverage visible. A diagnostic can have successful findings while another surface remains unread. The count concerns the successfully read population, not every possible area of the account.

Undefined ratios require their own labels. A missing denominator or no selected acquisitions should not become an ordinary numeric CPA or ROAS. A table can preserve the evidence without forcing a chart to imply a value.

The next step should target the gap. Resolve access, choose a compatible measure or ask the user for a required event. This makes the conversation useful without inventing a reassuring number or a dramatic unsupported explanation.

Write a useful first answer

Lead with the scoped result, then give the definition that makes it understandable. A count answer can name the selected population and current status context. A performance answer can state dates, source and chosen outcome before discussing its interpretation.

Use plain names and meaningful units. A client needs to recognise the campaign or group and understand whether the figure concerns money, sessions or selected conversions. Raw provider field syntax rarely improves the first response.

Keep observation separate from inference. A tool-returned decline in purchases is evidence of a measured movement. An explanation involving competition or website behaviour needs additional support. The answer should not turn a hypothesis into a confirmed cause.

Include a limitation only where it affects the conclusion. Missing coverage, an undefined ratio or an unconfirmed event matters. An implementation detail unrelated to the user's decision can stay behind the scenes.

Close with the next useful question or supported step. The user should understand whether the assistant needs a clarification, can retrieve a narrower view or can prepare a requested operation. A clear first answer makes follow-up work easier without forcing an unsolicited report or change.

Follow an invented agency conversation

The following scenario is invented and does not describe a customer account. Rowan Harbour Media asks how many campaigns are currently enabled in a connected client account. A successful complete selection returns one enabled campaign named Winter Services.

The reply states that scoped count and name. When the user says pause it, the assistant retains the resolved target and prepares a status-change proposal. It does not execute merely because the target is unambiguous.

The confirmation identifies Winter Services and the exact pause action. After the human confirms, the tool checks relevant current settings, sends the supported operation and verifies the resulting state where possible. The count read and live mutation remain distinct steps.

Later, the user asks for five possible headlines for the same offer. The assistant can draft them from supplied facts without an unnecessary account-performance or Keyword Planner read. The suggestions are not presented as tested winners.

If the user then asks to upload an agreed draft, the conversation moves back to a supported creation proposal with required parent and URL context. This example shows how a natural thread can contain reads, creative work and edits while preserving their separate evidence and approval requirements.

Keep follow-up instructions meaningful

A follow-up can refer to an established target, revise a measure or request another presentation. The assistant should retain unchanged context and apply the actual new instruction. Repeating the earlier selection after an explicit correction fails the purpose of a conversation.

For example, use all primary conversions changes an earlier purchase-only definition. Use total paid sessions changes a narrower traffic population. Show the same result weekly changes presentation grain while preserving the underlying period.

Resolve unclear pronouns only when necessary. If several candidates were listed, the user may need to identify which one. If one campaign was explicitly resolved and remains the topic, a generic request to choose a campaign again discards usable context.

Keep evidence age relevant. Previously returned settings can inform the conversation, while live execution rechecks the required current state. A stored plan is not permission to override a change another account user made in the meantime.

When the new question needs another dataset, explain the transition clearly. The assistant can preserve the existing scope while retrieving compatible new evidence rather than treating the follow-up as an unrelated request with default dates and metrics.

Avoid unnecessary reads

Start with local field and dependency information where it is sufficient. A catalogue can establish which relationship or action is supported without asking the provider for an account-wide inventory. Actual current settings need a read only when the question requires them.

Reuse returned data for requested formatting. A table or chart of an already retrieved compatible dataset should not trigger an identical second query. The framework memoises equivalent data plans within the request and can render from the returned data identifier.

A new metric, population or grain can require another read. Reuse should not mean inventing daily observations from a period total. Efficiency comes from avoiding redundant work while preserving the question's actual evidence requirements.

Do not fetch every source in anticipation of possible questions. If the request concerns a campaign status, a broad Analytics and Merchant read adds little. If the user asks about a product relationship, select the supported dataset and related evidence needed for that question.

Handle non-retryable plan failures deliberately. An unchanged invalid request does not become valid through repetition. Use its structured error to choose a supported adjustment or ask the user for the missing selection.

Answer creative questions conversationally

The agency should be able to brainstorm in the same thread it uses for account evidence. Suggestions for headlines, descriptions, names or keyword themes can use the product and offer context already supplied.

Ask for essential missing facts instead of inventing them. If the assistant does not know the product's features or intended audience, a short clarification is better than a confident draft containing unsupported guarantees or specifications.

Do not append unsolicited research exports or targeting boilerplate to an ordinary creative reply. The user may simply want a few useful ideas. Actual search volume, competition or bid-range questions are separate evidence-backed research requests.

Label the creative output honestly. It is suggested wording or an idea, not a Google-returned result, approved ad or proven winner. The user can discuss and revise it before requesting any account operation.

This flexibility matters to agencies because a client question often develops through discussion. The assistant can help think without forcing every sentence into an API workflow, while remaining strict about measured facts and live-change authority.

Use requested charts and tables

Render visuals when requested and from compatible returned data. A line can show movement over time; a bar can compare categories. A pie needs nonnegative additive parts of one measure rather than ratios such as CPA or ROAS.

Native conversational charts support bounded series and points. A long unqualified trend can use a coarser breakdown for readability. Explicit daily detail should remain daily within supported capacity, with any necessary split explained.

Preserve the selected period and labels. A formatting request should not shorten the data window or drop observations silently. Weekly labels should communicate dates rather than opaque provider codes.

Use compatible units on a shared axis and keep source series distinct. A comparison is not a combined total. Where values are missing or bounded, a table can preserve them more honestly than a chart that would imply a defined point.

Keep displayed metrics to at most two decimals and use understandable asset names. The output should help the client assess the answer, not expose technical identifiers or a large set of raw columns.

Choose a diagnostic workflow when needed

Some questions need a defined diagnostic rather than a performance table. An audit can inspect its current settings and structure checks. A policy read can inspect published policy findings. The Links workflow can read Google's destination evidence.

Keep their populations distinct. A policy reader can include non-removed paused ads, while an enabled-ad audit check has a narrower selection. Different totals can reflect scope rather than contradictory provider data.

The destination reader is not a live website crawler or form test. It reports published verdicts, with possible timing differences from the current page. A website owner may need to investigate the actual journey separately.

Audit heuristics also need business context. A missing negative list or low Quality Score can direct investigation without proving a quantified waste amount. The agency should not treat every checklist finding as a universal instruction to pause.

Use the appropriate related guides: agency audits, policy monitoring and destination checks. The first answer should identify which workflow supports the question.

Move from a question to an edit proposal

A read question does not authorise an account change. Once the user requests a supported operation, resolve the target and prepare the exact action. The assistant should preserve the relevant conversation context without skipping the human confirmation boundary.

Show recognisable asset and parent names. A keyword operation needs its group and campaign context; an asset-group operation needs the appropriate campaign. The proposal should not confuse a sibling with a parent or expose identifiers as the main explanation.

Separate logical actions clearly. Creating a group, adding keywords and enabling a target are different steps. The dependency framework can sequence required parents before children while avoiding unrelated lookups for operations that need no such dependency.

Respect actual support for each setting. The ordinary budget edit is for non-shared daily budgets. A readable provider field or audit recommendation does not prove that every possible write is exposed through the conversational tool.

The proposal states that no changes have been made yet and expires after its defined validity period. Confirm and Execute triggers the live edit. The user can review or correct the target and action while preparation remains reversible.

Explain execution evidence precisely

Before execution, the tool checks relevant authorisation and current settings. After the supported request, it verifies the resulting account state where possible. These checks protect the operation from stale assumptions in the original conversation.

An Executed and Verified result establishes the checked account change. It does not guarantee better performance, completed provider approval or a working website journey. Those questions need their own later evidence.

Uncertain execution should prompt inspection before retry. A request may have succeeded even if the follow-up read was incomplete. Blind repetition can create duplicate assets or another unintended change.

Distinguish an applied edit awaiting provider review from a confirmed rejection. They require different next steps. The reply should state the actual result instead of treating every unfinished milestone as a failed mutation.

Keep the subsequent business question defined. If the user asks whether the change helped, use an appropriate later period and selected outcome. A before-and-after observation can inform review without proving the edit caused every measured movement.

Transfer the question across colleagues

Record the client's question, account context, selected evidence and current answer. A colleague should not need to reconstruct a long thread to learn which conversion definition or campaign the result concerned.

Keep supplied business facts separate from retrieved metrics. The client can confirm an offer or service area, while the provider supplies the selected performance or settings. The handoff should preserve the provenance of each statement.

Identify the unresolved choice and next owner. A missing event correspondence, unclear target or unavailable read should remain visible. Another colleague can then continue with a concrete question rather than repeating a broad account review.

For an operation, record whether it was proposed, approved, executed or checked. These stages are not interchangeable. An uncertain result needs investigation rather than an automatic rerun.

Link the handoff to the relevant related workflow where useful. Client reporting concerns the decision conversation, while confirmation concerns the live operation. The account-question thread can connect them without merging their authority.

Set an accountable agency routine

A dependable routine helps staff ask precise questions without making the conversation unnatural. Start with the account, business purpose and relevant period; clarify consequential ambiguity; then use the smallest supported evidence plan.

Keep repeated questions consistent across clients while preserving their definitions. The same format can ask about purchases for one business and a named lead action for another. Consistency should not erase commercially important differences.

Use recurring reporting or monitoring for genuinely repeated evidence needs. A one-off account question does not automatically create a saved schedule. A notification also does not imply a colleague has reviewed or approved a change.

Review misunderstandings as definition problems first. When a client challenges a number, inspect the scope, dates and measure before repeating the answer. If the selection was wrong, correct both the result and any recommendation that depended on it.

Make the next step assessable. The agency can discuss ideas, retrieve facts and prepare supported edits while keeping their boundaries clear. This is what makes a general conversational framework useful across many account questions rather than a growing collection of narrow canned responses.

Turn a vague request into one answerable question

A request such as check the account can refer to several different jobs. The agency can ask whether the user wants current configuration, measured performance or a defined diagnostic. That clarification is useful because it changes the evidence plan. It should not become a demand for every possible setting before the assistant can help. Start with the consequential choice, then retain any facts the conversation already established.

For a current-structure question, identify the entity and status scope. How many enabled campaigns are there needs a current count, not a thirty-day performance table. For a performance question, identify the period and outcome. Which campaign has the lowest purchase-only ROAS during the selected dates needs the defined numerator, denominator and eligible population. These examples illustrate plan differences rather than promise that every wording resolves without clarification.

For a relationship question, specify the parent context and count unit. An excluded-product request can concern occurrences within enabled groups rather than unique inventory items. A follow-up about one campaign should preserve that campaign filter. The answer should not replace the user's intended unit with a simpler count merely because the preview is easy to inspect.

For a creative request, use the supplied offer and requested constraints. A user asking for five headlines may specify tone, length or claims to avoid. These instructions can shape the draft without account metrics. If the user asks which version performed best, the task changes to an evidence question and needs the relevant existing assets, period and selected measure.

Keep a corrected question connected to the corrected result

When the user says that the assistant chose the wrong population, the next response should make the changed selection clear. A correction from one paid source to all paid traffic is an analytical change, not merely a new title. The assistant should use a compatible plan for the requested population and avoid presenting the old result under the revised label.

If the original response included a recommendation, assess whether that recommendation still follows. A ranking based on all primary conversions may differ from a purchase-only ranking. Correcting the metric while leaving an earlier pause recommendation untouched can carry the original error into a live proposal. The agency should review the recommendation against the corrected evidence before requesting an operation.

Preserve the rest of the context that the user did not change. A corrected metric does not necessarily imply new dates, another account or a different campaign. This prevents clarification from resetting the whole conversation and makes the response easier to compare with the user's original intent.

Record the revised question and its successful evidence definition in the handoff. A colleague should be able to distinguish the superseded answer from the current result without reading every exchange. This is an operating practice, not a new automatic ticketing feature. It keeps a useful conversational correction from disappearing when another person prepares the next report or proposal.

Check the human-facing wording after a corrected read. The total, table title, chart series and next-step explanation should all refer to the same revised definition. A correct internal selection can still confuse the client if the reply retains an old purchase label or period heading. The agency should assess the full output as one answer, with the corrected scope visible wherever it changes the meaning.

For a consequential change request, resolve any remaining ambiguity before the confirmation is prepared. A corrected report can inform the decision, but it does not authorise a mutation or prove which target the user wants changed.

Agency account-question FAQ

Does every question need a provider call?

No. Creative suggestions and general explanations can use supplied context. Actual account numbers, settings and existing assets require the appropriate returned evidence.

Can a follow-up refer to the same campaign?

Yes, when the target is resolved unambiguously in the usable context. Ambiguous references need clarification, but established context should not be discarded without a reason.

Can the assistant compare a named conversion with an event?

Use a resolved Ads action and an Analytics event the user names or confirms. A suggested event name alone is not confirmation of the business correspondence.

Does a count mean the displayed examples?

No. Counts should use a complete supported scoped selection. A bounded preview is for readability and should not become the population total.

Can a question trigger a live edit?

A requested supported action can lead to a proposal. The live mutation occurs only after the human uses Confirm and Execute, with the exact target and operation visible.