What predicted conversion rate measures
When someone searches, the auction happens in the moment, and the bidding system has to decide what that specific click is worth before anyone has clicked anything. Predicted conversion rate is its answer: an estimate of how likely this particular person, on this device, at this time, with this wording, is to do the thing you are counting as a conversion.
The estimate is built from signals available at auction time — the query itself, device, location, time of day, language, browser, and whether the person sits on any of your audience lists — combined with what the account’s own history says about similar situations. Multiply that estimate by the value or target you have set, and you have the bid. This is the whole mechanism of Smart Bidding in one sentence.
Why predicted conversion rate matters
It explains behaviour that otherwise looks arbitrary. Why two people searching the same phrase see your ad in different positions, why bids rise on weekday mornings and fall late at night, why a returning visitor is bid on more aggressively, why costs move after a landing page change — all of it comes back to the same estimate shifting.
It also explains why data quality matters more than settings. The prediction is only as good as the conversions it learned from. Feed it a soft goal such as a page view and it becomes very good at predicting page views. Feed it duplicated conversion actions and it learns from double-counted history. Nothing in the interface will tell you the model is confidently wrong; the account simply spends against a poor definition of success.
Where it goes wrong
The first misunderstanding is confusing it with the conversion rate in your reports. Reported conversion rate is a historical average across everything that happened. Predicted conversion rate is a forward-looking estimate for one auction, it is never shown to you, and no column contains it. Comparing them is not possible and not useful.
The second is disturbing the model constantly. Every change to conversion actions, attribution, targeting or bid strategy resets some of what it had learned, and an account edited every few days never gets past that. The third is thin evidence: with few conversions the estimate leans heavily on broad patterns rather than your own account, which is a real limitation for advertisers in smaller markets, where a campaign may simply not produce enough weekly conversions for the model to become confident.
What to do about it
Improve the inputs rather than hunting for the number. Count one clear primary conversion that represents a genuine business outcome, remove duplicates, and where sales close offline import the result so the model learns which clicks became customers rather than which became enquiries. Upload and refresh customer lists, because a returning buyer is a signal the model can use.
Then leave it alone long enough to learn. Make one change at a time, wait out the learning period, and judge the result on cost per acquisition rather than on daily movement. If the account genuinely cannot produce steady conversions, that is an argument for a simpler bid strategy and better tracking first, not for more automation — a point worth settling early in any ongoing account management.