Build Context Hints From Sales Calls, Not From Brainstorms
By Keshav Parsai 6 min read
In brief
The best context hints already exist in your sales call recordings and closed-lost notes, written in the buyer's language rather than the marketer's.
Last verified: 12 September 2026 | Version: 1.0 | Next scheduled review: 12 December 2026
Most context hints are written in a meeting. Someone opens a document, the team describes their ideal customer, and the result is a set of hints that describe how the company sees its market.
The matcher does not work on how you see your market. It works on the meaning of what a real person typed, unprompted, while trying to solve something. The closest available record of that language is not in your marketing. It is in your sales calls, your support tickets and your closed-lost notes.
Why invented hints underperform
An invented hint describes a category. A real buying situation has a trigger, a constraint and usually a failed alternative, and the invented version almost never contains any of the three.
Compare:
Invented: "Companies looking for better expense management software."
From a discovery call: "Our finance person is spending the last three days of every month chasing receipts over Slack, and we just failed an audit sample because two of them were missing."
The second one names when this becomes urgent, who feels it, what broke, and what the buyer is actually afraid of. Every one of those is match surface you cannot invent, because you were not there.
The extraction method
This takes an afternoon and produces more usable hints than a quarter of brainstorming.
One. Pull thirty conversations. Ten closed-won, ten closed-lost, ten disqualified. The lost and disqualified ones matter as much as the wins, because they tell you which situations to keep out of your hints.
Two. Pull only four things from each. Do not summarise the call. Extract the trigger event in the buyer's words, the alternative they were using, the constraint that ruled options out, and the phrase they used for the problem. Four lines per conversation, 120 lines total.
Three. Cluster by trigger, not by feature. This is the step that changes the output. Grouping by feature reproduces your product marketing. Grouping by trigger produces buying situations, which is what a hint describes. "Just hired employee number three", "failed an audit", "the person who ran this left" are triggers. Each cluster becomes a candidate ad group.
Four. Keep the buyer's noun. If eleven calls said "receipt chasing" and your website says "expense capture automation", the hint says receipt chasing. Your product name goes in the ad, not the hint.
Five. Write one hint per angle, one theme per ad group. Three to six hints describing the same trigger from different directions, as covered in our piece on hint count.
Six. Mark the disqualifiers. Every closed-lost reason is a situation to leave undescribed. There is no exclusion layer on this channel, so the only exclusion available at the hint stage is silence.
What the closed-lost pile gives you that nothing else does
Wins tell you which situations to describe. Losses tell you which ones will cost you money.
If eight of ten losses were companies under fifteen people who wanted a free tool, that is not a reason to write a hint about small companies. It is a reason to leave company size out of your hints entirely and put a price qualifier in the ad copy, so the same reader self-selects out before you pay for the click.
This is the part a brainstorm structurally cannot produce, because nobody brainstorms the customers they do not want.
For companies with no sales calls yet
The same method runs on weaker inputs, in descending order of value: support tickets, sales email threads, the free-text field on your demo form, review sites in your category, and the question threads on Reddit or industry forums where people describe the problem before any vendor is involved.
The one thing not to substitute is competitor marketing copy. It is the invented version of somebody else's hints.
The limit of this method
Mined hints are grounded in past buyers, and past buyers are a biased sample. They are the people your previous channels reached. If your Google Ads and outbound have been finding mid-market companies, your call recordings describe mid-market situations, and hints built from them will keep you in the same segment.
That is fine when the goal is more of what works. It is a real problem when the reason you are testing this channel is that ChatGPT reaches buyers your existing channels miss, which is one of the more credible arguments for the channel in the first place.
The correction is to run one deliberately un-mined ad group alongside the mined ones, built from a situation you believe exists but have never sold into. Keep it separate so its results stay readable, and treat it as the exploration budget.
What we cannot tell you
- Whether mined hints outperform invented ones. No published test exists and we have not run one.
- How many conversations are enough. Thirty is a practical sample size, not a validated threshold.
- Whether the matcher favours buyer language over marketing language. Undocumented. The argument above rests on meaning matching, not on evidence about vocabulary.
- Which mined hint performed. No hint-level reporting exists.
Quick answers
Where do good context hints come from? Recorded sales calls, support tickets and closed-lost notes, where buyers describe their situation in their own words before a vendor reframes it.
What should I extract from a sales call for this? Four things: the trigger event, the alternative they were using, the constraint that ruled options out, and the phrase they used for the problem.
Should I group hints by product feature? No. Group by trigger event. Feature grouping reproduces your product marketing rather than a buying situation.
What do I do with closed-lost reasons? Leave those situations out of your hints, and put the disqualifying condition into your ad copy instead. There is no exclusion layer to block them with.
What if we have no sales calls yet? Use support tickets, demo form free text, and forum threads where people describe the problem before vendors appear. Do not use competitor copy.
Sources
| Claim | Source | Tier |
|---|---|---|
| Hints should describe genuine use cases in clear natural phrases | OpenAI Help Center, Create Ad Groups for ChatGPT Ads, August 2026 | Confirmed, primary |
| Hints are matched on meaning of a live conversation | OpenAI Ads Manager documentation, 2026 | Confirmed, primary |
| No exclusion layer exists | OpenAI documentation, by absence | Absent |
| Build hint themes from actual buyer questions | Lapis, context hints best practices, 2026 | Reported, third party |
| The extraction method and the sample bias correction | Newtation method | Ours |
Related reading
- How to Write a ChatGPT Ads Context Hint
- How Many Context Hints Should One Ad Group Have?
- Can You Exclude Conversations on ChatGPT Ads?
- Context Hints for Long B2B Cycles
Changelog
12 September 2026, v1.0. First publication. Documents the six-step extraction method and the past-buyer sample bias it carries.
Field kit
Tools
Site
OpenAI (primary)
Keshav studies how AI systems retrieve, verify, and cite brand information. At InPromptAds, he leads source research and turns platform documentation into practical guidance for advertisers.
