# Bid strategy selector — research and method Last reviewed 2026-09-29. ## The governing refusal A bidding strategy is a set of rules, and a rule does not forecast a market. So this tool produces no predicted cost per click, no predicted clicks, no predicted conversions and no predicted return. The refusal occupies the third preview statistic, permanently reading "Not predicted", so the gap is visible exactly when someone is looking for a number to fill it. The second refusal is narrower and easy to get wrong. A target CPA or target ROAS is a *constraint on bidding behaviour*, not a set outcome. It does not set the campaign's average cost and it does not cap any individual bid. It changes which auctions the system finds worth entering aggressively, and the volume consequence is the entire point. So the tool explains the constraint mechanism rather than treating the target as a promise, and it never recommends a target figure: a sensible target is a function of the operator's margin, price and conversion lag, none of which this page can read. Any number produced here would be a round figure from nowhere with a spreadsheet's credibility. ## The ordering decision, which is ours The page states a fixed order for the eligibility checks, and that order is judgement, not a platform fact: 1. Can the account measure a conversion at all 2. Is there enough conversion history to learn from 3. Is there a value, or only a count 4. The stated goal 5. How much control the operator wants to keep Control comes last for a specific reason: it is the only step that cannot rescue an unready account. Applying it early is what produces a campaign on a target strategy, optimising a misconfigured conversion, with nobody able to tell for several weeks — because the automation will not report a problem, since from its point of view the tracking is working perfectly. The order runs the other way too, legitimately, and the tool supports it: if the operator is buying data in a new market, or optimising for something conversion tracking cannot represent such as an unlogged phone call, a return target is the wrong instrument and Manual CPC is the right one. The goal question exists so that choice can be made deliberately. ## Asserted as documented platform behaviour **The strategy set, and what each optimises for.** Manual CPC — you set every bid. Maximise Clicks — the most clicks within budget, optionally under a maximum CPC limit. Maximize Conversions — the most conversions within budget, with the average cost reported as an outcome. Target CPA — Maximize Conversions with a target cost as a constraint. Maximize Conversion Value — the most total conversion value. Target ROAS — Maximize Conversion Value with a target ratio as a constraint. Enhanced CPC — bids for clicks, raising the bid on clicks predicted to convert, documented as being retired in favour of the others. **The requirements.** The conversion-based strategies require conversion tracking and a history of conversions. Maximize Conversion Value and Target ROAS additionally require conversion values. Maximise Clicks requires no conversion data. Manual CPC requires nothing. **Conversion values must be real numbers.** Uniform placeholder values make a value-based strategy behave identically to a count-based one, so it provides no advantage while appearing to. **Enhanced CPC is documented as being retired.** It is included for accounts still running it, and is not pickable. ## Asserted as our own judgement, and labelled as such - **That the checks run in the order above.** Opinion, stated as such. - **That partial conversion tracking produces caution rather than eligibility.** A partial sample is usually biased in a structural way — the recorded conversions tend to be the ones that happen online, immediately, without a phone call — and an automated strategy will find that pattern. Google documents a requirement for conversion tracking; it does not document that a partial implementation is equivalent to a complete one. - **The conversion-history bands.** Google requires a history for the conversion-based strategies and publishes no count that unlocks them. The bands are practical guidance, and the page says so both where they appear and in the sources section. Presenting a rule Google does not publish would be the exact failure this project is built against. - **That a tight target should be loosened rather than tightened into place.** Standard practical advice, offered as advice. - **That Target ROAS on a small account is the most common misapplication in the category.** The mechanism is documented — the strategy optimises a ratio, so it protects that ratio by declining searches it cannot predict — but the frequency claim is our observation and is not asserted as Google's. - **That Manual CPC is a legitimate choice rather than a fallback.** A framing decision, and the opposite framing is what makes the tool useless on an account whose tracking is broken. The page separates the two explicitly in the sources section. ## A distinction the page is careful not to blur A maximum CPC limit and a target CPA are both called "limits" in ordinary speech, and they are not the same kind of thing. A maximum CPC limit is a genuine ceiling: Google will not bid above it for an individual click. A target CPA is not a ceiling on anything. It constrains average bidding behaviour across the campaign, and a week is a short window against a learning system. The tool draws the two with different edges and explains the difference, because grouping them would transfer a stronger guarantee to the weaker instrument. ## Two axes, and why an option was removed from one of them The form asks what you are optimising for, and separately how much bid control you want to keep. The 1.0.0 build also offered "Maximum control over every bid" as a *sixth objective*, which made the two questions overlapping and created a silent failure: the objective ladder had no branch for that value, so it reached the final branch, which assumes the objective is Target ROAS. An operator who selected the one option that described their actual objective, then declined maximum control in the next question, received a recommendation to set a return floor. The option is removed rather than given a branch. Two answers to the same question is the underlying defect; adding a branch would make the page produce a sensible answer while leaving the form asking something twice. The objective question is now the objective, and the control question is the control. Two rules follow from this, and both are asserted by tests rather than left to review: - **The control axis is evaluated before the objective and wins.** With maximum control selected, Manual CPC is the answer and the objective is never read. - **An unrecognised objective value never falls through to another branch.** The final branch is now explicitly the Target ROAS branch, and an unknown value gets a fallback that says the page could not read the question. A ladder whose default case is also a real case will eventually answer one question with another question's answer. ## A requested limit against a target-less strategy is a mismatch Maximize Conversions and Maximize Conversion Value run without a target. If the operator has asked to "automate, with a limit", the recommendation they receive may still be one of those, because sometimes the volume is what the account needs. That combination is a **mismatch between what was asked for and what the recommended strategy can do**, and it is reported as its own state. It is not described as a target being in play: no target exists, and saying so in the warn colour would present the recommended setting as a fault. Nor is it silently resolved in either direction. The page names the mismatch and offers the honest alternatives — the budget becomes the only ceiling, or move to a cost target once there is enough history for the target to mean something — while acknowledging that a maximum CPC limit is a third option and a weaker guarantee, because it caps a click rather than a conversion. ## Not asserted anywhere - Any predicted cost per click, clicks, conversions, cost per acquisition or return. See the governing refusal above. - Any recommended target figure, including as a worked example. A number in an example is a number people use. - Any claim that a given account is or is not ready. The tool never sees the account; it takes the operator's answers as given and cannot detect a mis-transcribed one. - Any Ad Rank figure. - Any claim about how the bidding system weights a target internally, or how quickly it settles after a change. - Any conversion count that Google does not publish as a threshold. - Any statement that a specific strategy is what Google recommends. That recommendation changes, and the page says it has a shelf life. ## Known limits - The tool takes the five answers as the operator reports them. It cannot read the account, so it cannot detect a mis-transcribed answer, and a wrong "yes" to conversion tracking produces a confidently wrong recommendation. - The history bands are self-reported and self-bucketed. The page asks for a band rather than a count partly for this reason, and partly because a count invites a precision the underlying guidance does not support. - A strategy marked with a caveat is not blocked, and the recommendation may still be that strategy. Partial tracking is treated as something to fix first, and the first step says so, but a caveat is not a prohibition. - The tool will return something other than the operator's stated goal when a prerequisite is unmet. The reasoning text says so explicitly, but a user who asked for Target ROAS and was handed Maximize Conversions may not read it. - Google changes its bidding options. The page tells users their account is authoritative and that if it disagrees with the page, the account is right. ## Verification `tests/run.mjs` checks that preference never overrides an unmet prerequisite, that the three eligibility states stay distinct, that each unmet requirement is named rather than shown as an unexplained block, that the refusal holds across several states, that no currency figure appears in the rendered results or the exported CSV, that no target figure is produced, that the maximum CPC limit and the target CPA are described as different kinds of thing, that a clean account produces no manufactured prerequisite work, and that the CSV export carries the refusal with it.