> ## Documentation Index > Fetch the complete documentation index at: https://docs.erdo.ai/llms.txt > Use this file to discover all available pages before exploring further. # Growth reviews > Turn research and campaign evidence into owned investigations, experiments and measured decisions. A growth review uses an existing [Workstream](/workstreams.md) to turn evidence into work: investigate a finding, prepare a candidate, propose a change, or evaluate an [Experiment](/experiments.md). It uses the workstream's current authority and budget. Ad spend and publication remain separate approval decisions. ## Seed the review with research Ask your agent to save a reviewed study as workspace-only [Knowledge](/knowledge.md) and attach it as an input to the campaign workstream. Include the sources and their dates, the account and market covered, the observation window, a short ranked set of hypotheses, and what remains unknown. A notebook on your computer is not automatically available to the recurring agent. Deep exploratory analysis can remain outside the recurring review. Save the findings it should use and the questions that need further investigation. Research estimates suggest tests; they do not establish conversion improvements or authorize a campaign change. ## Review a bounded slice of current evidence The Lead Engine method reviews acquisition opportunities weekly by default, or when material new evidence arrives. The workstream's research input can set another cadence. Daily health checks continue on their existing schedules. The review considers the actual search queries and keywords buying traffic, their spend and captured enquiries, offer fit, demand that current coverage may miss, and the steps from a page visit through form completion. Named competitors are candidate acquisition audiences, not automatically irrelevant traffic. Counts use completed dates in the ad account's time zone. Repeated rolling windows overlap and cannot be added as separate weeks. Provider conversions, captured enquiries, qualification and reachability remain separate outcomes; missing or delayed evidence stays unknown. Preserve an approved, successful customer exception when reviewing or repairing a live development. Contact-first is the default for new development work, not permission to overwrite an existing form contract. Audit the actual first save and final save, the desktop and mobile ad destinations (including paused states), the pipeline schema and transform, and the consent fields before proposing a change. Treat pipeline requiredness, rendered requiredness and fields present in the first save as separate observations. A permissive capture schema supports partial contacts; it does not establish which fields the visitor must answer. For a top-of-funnel diagnosis, compare actual daily budget, spend, CPC, clicks, serving status and any returned impression-share loss fields. Empty provider fields remain unknown; budget constraint and low demand are different findings. When page evidence is available, pair pageview → CTA/modal → `form_start` → capture → canonical enquiry, using the same geo, device, page, attribution scope and timezone. Modal opens and modal submits are distinct events. Verify what each event means: the landing kit emits `form_start` on first field focus. A capture probe checks the capture path; it does not measure whether visitors want the offer. For a lead-form diagnosis, use the same-page, same-date/session funnel `pageview → modal_open → form_start` (first field focus) → first successful capture. Keep failed submissions separate from low opener/start rates, and keep commercial qualification separate from capture. Do not treat every form start as a submit, compare rental and luxury benchmarks as causal evidence, or call unknown search-term coverage waste. Before a price increase, distinguish low demand within the current targeting (few impressions despite high impression share and little share lost to budget or rank) from rank or quality loss; poor Quality Score is diagnostic, not an auction input. Read the relevant ad copy, search terms, offer fit and competitor intent, preserving competitor traffic as a distinct hypothesis rather than discarding it. Missing evidence remains unknown, and the existing quality rules and thresholds remain the controlling method. The same review separates contact saved from qualified or reachable. Do not pause a converting generic-keyword destination solely because Quality Score is low or the label is generic: quantify spend, captures, downstream qualification, reachability and lost volume, then use a reversible scoped experiment when evidence is thin. Audit all public destinations, including mobile and paused states, but do not launch paused campaigns or change budgets during the audit. On the contact step, the kit's omitted required flag defaults to `true`. When `stepIndex > 0`, an omitted or explicitly false required flag is optional, while an explicit `required: true` remains required. Read the rendered configuration rather than inferring requiredness from a shared component or tolerant schema. ## Follow the work through The agent first checks existing assignments, experiments and pending decisions. A finding already owned by one of them advances there rather than becoming a second recommendation. New findings can lead to: * A bounded assignment to investigate evidence or prepare a private candidate. * An experiment with a hypothesis, exact treatment, metrics, review window and conditions for stopping or continuing. * An approval card describing a concrete live change and its evidence. * One choice or escalation when a decision genuinely needs a person. A pending choice blocks the action it governs. Other investigation and private preparation can continue within the existing authority. Holding an active test steady protects its measurement window; it does not prevent learning about the rest of the funnel. ## Read results honestly Page-bound experiments have automatic measurement setup. Other treatments, such as a keyword expansion, need explicit configuration of the existing deterministic measurement recipe and a check that observations reach the experiment ledger. Page traffic-allocation policy does not apply to those tests. A change that was already applied without a declared control remains an observational study when recorded later. Its before-and-after results can inform a decision, but do not prove that the change caused the difference. If enquiry attribution cannot identify a keyword, the review reports that gap rather than assigning the enquiry by guesswork. A material update produces a workstream digest explaining what is being done, why, the next owned step and what evidence would change the decision. A review can recommend improving or expanding acquisition as well as reducing waste; it does not need to make a change on every pass. Unchanged checks and waiting on an existing decision stay quiet; explicitly requested scheduled briefs retain their requested cadence. Any scaling proposal states the budget change, duration, maximum incremental spend, including provider overdelivery and limit enforcement, review point, stop conditions and existing approval authority. It does not rely on a universal conversion-count threshold. Interventions and confounding factors stay with the existing observational window and owner. ## Reuse a reviewed method The built-in **Research — Apply reviewed findings to growth** skill (`research-to-growth`) connects a reviewed research input to the existing review. Lead Engine setup, campaign planning, recurring reviews and Growth Engineer use the same method. It does not create another scheduler. The agent chooses the method that fits the question: * **Comparable developments:** research sourced property facts, then run the attached **Research — Property comparable screening** calculation. Distance, asking-price overlap and sale stage help shortlist competitor keyword candidates. Local assumptions must be explicit; incomplete comparisons stay unknown. Similarity is a research heuristic, not predicted conversions or a bid. * **Locality and country/language demand:** corroborate relevant search demand, keyword coverage and offer fit, then verify that the page and follow-up support the proposed audience. A new development can use this recipe without borrowing another development's conversion rates. * **Existing keywords and funnel steps:** calculate counts, rates and costs from the owner's scoped evidence, retaining unknown attribution and unanswered qualification. Investigate the step that limits progress; expansion, content, capture, qualification and reachability can all be subjects of owned work. Skills describe when and how to use a method. Attached scripts perform repeatable arithmetic through the existing calculation tools. The agent interprets the result and prepares a proposal; existing authority determines whether it can act. No separate model-serving system is required for these methods. A fitted predictor or allocation model needs its reviewed executable, fitted artifact, feature definitions and applicability before a recurring agent can use it as a model. A lab scenario can inform a hypothesis with its limitations; the agent must not invent coefficients or retrain a model during a routine review. The handoff retains the method version, research/source revision, input snapshot, parameters, scope and evidence dates alongside the resulting Assignment or Experiment. Measurement and the next decision close the local loop. A proposed change to the shared method returns to human review; it does not automatically rewrite instructions for every workspace. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.