Ask ChatGPT for Google Ads keywords and it hands you fifty in ten seconds, complete with a search volume and a suggested bid for each. The list looks finished. The numbers are fiction — ChatGPT has no access to Google's search data, so every volume and every CPC in that table was made up to look plausible. Keyword research with ChatGPT works, but only when the model does the part it is good at and the numbers come from somewhere real.
Why ChatGPT alone gets keyword research wrong
A language model is trained to produce likely text, and "dog groomer near me — 2,400 searches a month — $3.20 CPC" is very likely text. It is not data. Ask the same question twice and the volumes change; ask about a niche the model has barely seen and the confidence stays exactly the same. A keyword plan built on numbers ChatGPT produced from thin air is a plan built on noise.
What the model is genuinely good at is language. It expands one seed into thirty phrasings a real customer might type. It sorts five hundred queries into intents in seconds. It notices that every query that converts for you shares a modifier no one on the team had named. Those are exactly the parts of keyword research a human finds slow.
So the division of labour is simple. Real numbers come from two places: Google's Keyword Planner, which gives volume ranges and bid estimates, and your own account, which gives the only numbers that matter — what each search actually cost you and which of them actually converted. ChatGPT's job is everything between the numbers.
The best keyword list is already in your account
Keyword Planner tells you what people search. Your search terms report tells you what people searched right before they paid you. Somewhere in that report are queries that converted even though they match none of your keywords — broad match or a close variant dragged them in. Each one is proven demand, already matched to a proven ad and a proven landing page. The only thing missing is the keyword.
That is also the one keyword source your competitors cannot copy. Everyone sees the same Keyword Planner output; nobody else sees your search terms report.
Finding those queries by hand means an export, a pivot table, and a cross-reference against your active keyword list — an hour in a spreadsheet. It is one message once you connect ChatGPT to Google Ads; the same works with Claude.
Prompt 1: the searches that convert but are not keywords
Pull my Google Ads search terms report for the last 90 days. Find every search term that converted at least once but does not match any keyword I am bidding on. For each one show the term, its campaign and ad group, spend, conversions and cost per conversion. Tell me which ones deserve to become keywords, and for those propose the exact keyword, the match type, and the ad group it belongs in. Do not apply anything.
Ninety days, not thirty: conversions are sparser than waste, and a converting query that appears once a fortnight still deserves the promotion. What the promotion buys you is control. As its own exact-match keyword, the term gets its own bid instead of whatever broad match felt like, it serves reliably instead of when the auction happens to wander your way, and it starts accumulating query-level history you can act on.
The same report has a second half — the terms that spent and never converted. That is a different job with a different filter, covered in the wasted-spend walkthrough. This prompt adds; that one subtracts. Run both.
Prompt 2: grade the keywords you already pay for
New keywords compete for budget with old ones, so research includes an honest look at the incumbents.
List every keyword in my Google Ads account with its match type, status, spend, conversions and cost per conversion for the last 90 days. Flag three groups: keywords with meaningful spend and zero conversions, keywords whose cost per conversion is more than double the account average, and keywords with almost no impressions. For each flagged keyword recommend keep, pause, or change match type, with one sentence of reasoning. Do not change anything.
The first two groups fund the new list: every dollar a dead keyword stops spending is a dollar a proven query can have. The third group is diagnostic rather than financial — a keyword with no impressions is either too low in volume to matter, scored too poorly to serve, or silently blocked by one of your own negatives. That last case is more common than most accounts think, because a conflicting negative keyword wins without any warning.
Prompt 3: expand into new themes — and validate before you spend
This is the use everyone starts with — "give me keyword ideas" — done in the order that makes it safe.
Here is what I sell: [product, price point, who buys it]. Based on my converting search terms from the last 90 days, which intents and themes am I not covering yet? Propose 20 candidate keywords grouped by intent — comparison, problem-aware, solution-aware, brand-adjacent — and mark the group you would test first. Do not invent search volumes; I will check volumes in Keyword Planner myself.
Grounding the request in your own converting terms is what separates this from the generic fifty-keyword list: the expansions extend what already works instead of describing your industry from orbit. The last line matters just as much — say out loud that invented volumes are not welcome, and the model stops offering them.
Then the loop closes outside the chat. ChatGPT proposes, Keyword Planner confirms the demand exists, and two weeks of live serving tells you what neither could: whether the clicks are worth the money.
Where a new keyword should start
Match type at launch should follow the evidence behind the keyword, not a house style.
| The keyword came from | Start it on |
|---|---|
| A query that converted in your search terms report | Exact |
| A theme with some conversions behind it | Phrase |
| A hypothesis from Prompt 3 | Phrase, with a modest bid |
Broad match is a discovery engine, not a starting point: it only behaves in campaigns with smart bidding and steady conversion volume, and even there it needs a well-tended negative list as a fence. A keyword promoted from your own converting searches needs no faith at all — exact match simply formalises something the account already proved.
From a list to a live campaign
Where the new keywords land matters as much as what they are. A term that shares intent with an existing ad group joins it. A genuinely new theme gets a new ad group, so it can earn ads written for it instead of borrowing someone else's. A new offer entirely is a campaign — here is what that looks like from inside a chat, including the four defaults the create step leaves for you to decide.
However the list goes live, the approval model is the same as everywhere else: the assistant shows the exact keywords, match types and destinations it is about to add, and writes nothing until you confirm.
Keyword research is not a project, it is maintenance — queries drift with seasons and competitors, and the report refills. The weekly Google Ads review keeps Prompt 1 in the rotation, and the full prompt library covers the jobs beyond keywords. In a vertical account the drift has a shape: an auto repair shop refills with part names and how-to searches, which is the subtracting half of the same report. In a landscaping account the drift is seasonal as well as thematic, which is why that one reads ninety days rather than thirty. In a moving company the report refills with destinations, and a destination that converted without ever being a keyword is how a new lane announces itself.
