Four fictional drafts for a scheduling service, perhaps only two meanings:
Scheduling appointments for visiting technicians.
Appointment planning for technicians who visit customers.
Managing customer visits for service technicians.
Reassigning a technician visit when the original assignee is unavailable.
The first three may be different ways of describing the same task. The fourth introduces reassignment, which could depend on a separate capability. A literal duplicate detector would keep all four because the strings differ. An editor needs to decide whether each line adds useful context to the ChatGPT ad group.
OpenAI says context hints supply information about the ads and are not exact-match keywords. Its guidance does not establish that adding synonymous phrases buys more coverage. Review the list for meaning and accuracy instead of treating it like a collection of search-query variations. Official targeting documentation
Put related lines beside each other
Copy the working list into a review document and give every row a stable reference. Keep the original version so you can understand what changed. Sort candidates into small groups by the customer task they describe: arranging a visit, changing an assignment, notifying a customer or checking completion.
Do not sort only by the first noun. “Technician” can appear in several genuinely different tasks, while “field worker” and “service engineer” may describe the same one. The task verb and practical constraint are often more useful for deciding whether two rows overlap.
For each group, write a plain-language summary of its distinct meaning. If you cannot explain the difference between two summaries, investigate whether the groups should be combined. If the distinction depends on a feature the offer lacks, remove that unsupported candidate rather than preserving it for variety.
Choose a representative, not a compromise sentence
When several hints mean the same thing, choose the clearest supported wording. Avoid stitching every phrase into one long sentence. “Appointment planning, scheduling and managing visits for technicians, engineers and service workers” is longer without necessarily adding information.
The representative should make the use case understandable without requiring the reviewer to decode your internal terminology. If customers consistently say “service visit” and the product team says “field intervention”, choose the term that accurately communicates the task to the intended reader of the brief. Keep a glossary note if the terms have different meanings in your business.
Do not select a phrase merely because it sounds more enthusiastic. “Effortless technician scheduling” adds an ease claim that the plainer alternatives did not contain. Deduplication should not introduce a new promise while removing repeated wording.
When two similar hints should remain
Similarity is a prompt to inspect, not an automatic deletion rule. Consider these fictional candidates:
| Candidate A | Candidate B | Reason to investigate separately |
|---|---|---|
| Schedule a first installation visit | Schedule recurring maintenance visits | Recurrence may require a different workflow |
| Share a visit window with a customer | Notify a customer when a visit changes | Initial confirmation and change notification are different tasks |
| Allocate work to an available technician | Reassign work after an absence | Reassignment may involve constraints not covered by initial allocation |
Retain both only if the advertised offer supports both. Similar wording can hide an important requirement, while very different wording can conceal the same meaning. That is why a simple word-overlap score cannot finish the editorial decision for you.
Similar vocabulary, different tasks
- Schedule a visit
Describes the initial assignment of an appointment.
- Plan a visit time
May have the same meaning. Compare before retaining both.
Reconsider - Reassign after absence
Describes a different event needing its own product support.
Checkpoint
Distinguish a duplicate from an umbrella phrase
“Managing service visits” may cover both booking and changes, while “booking the first visit” describes only part of that work. These are not pure synonyms. Ask what the broader row adds beyond the concrete tasks. If nobody can name it, the row may be redundant. If it represents another workflow step, describe and verify that step directly.
Do not automatically choose the broadest sentence as the representative. That can shorten a list by removing useful information. In the fictional scheduling example, two clear rows about booking and reassignment are easier to review than one row promising “all visit management”. The decision concerns the list’s meaning, not the number of imagined search phrases it might cover.
Record what was removed and why
A useful change note is short and specific: “Rows 4 and 9 merged into row 2 because all describe initial appointment scheduling.” Another might say, “Retained the reassignment row because it describes handling an unavailable assignee, confirmed in the current workflow.” These notes let another person review the decisions without reconstructing the entire exercise.
Keep removed candidates outside the list intended for use. If a line is waiting for product verification, mark it as unresolved in the working document. Do not mix it with approved wording and expect someone later to remember the distinction.
There is no universal target count in this process. A list with fewer rows can still be incomplete; a longer list can contain valid distinct tasks. Stop when every retained row has an explainable role and every relevant supported task in the brief is represented without unnecessary repetition.
Save the intended list, then check it
If the review leads to an API update, remember the documented update behaviour: supplying context_hints replaces the current list. Read the existing state and build the complete list you intend to retain before submitting an edit. Afterwards, retrieve the ad group and confirm the saved result. Omitting the field leaves it unchanged; an empty array clears it. The official source above documents these behaviours.
The editorial diff and the saved-state check answer different questions. The diff explains why a meaning was removed or preserved. The retrieval confirms what was actually stored. Neither confirms that a particular conversation will trigger delivery.
If disagreement is really about how much detail a phrase needs, use the specificity review. If nobody can support the distinction with a product fact, use the evidence check. Deduplication is complete when another reviewer can read the remaining list and explain what each line adds.
Sources and scope
Removing redundant meanings from a prepared hint list while preserving distinct use cases.
Working methods and examples are editorial suggestions. Check current platform requirements and available features before implementation.