1. Compile a brief before generating
A good naming brief states what the product does, who it serves, the painful job, the outcome, the mechanism, the reason to switch, important channels, desired tone, future scope, domain constraints, and names or roots to avoid. Missing information should stay visible rather than being invented by the model.
One sentence can be enough to begin. The important step is to turn the sentence, README, or product prompt into an editable record that a founder can correct before generation. This prevents polished names from being optimized against a misunderstood product.
2. Generate breadth in distinct directions
A list of 100 names is not broad when 80 are the same root with different suffixes. Generate in named veins: strategic benefit, mechanism or category, metaphor and imagery, and coined or phonetic forms. Keep the vein and rationale with each candidate so a strong direction can be expanded deliberately.
Generation should be allowed to fail transparently. If a hosted model is unavailable or returns an incomplete structure, a deterministic starting set can preserve the workflow. It should be labeled as fallback rather than passed off as model output.
3. Separate quality from availability
A weak name with a likely-unregistered domain is still weak. A strong name with a registered exact .com may still be worth comparing if another acceptable launch path exists. Score usability dimensions separately from registry and namespace evidence, then let the brief define which availability rule matters. When the field is large, run the survivors through the bulk domain checker so registry evidence arrives with sources and timestamps instead of a bare green badge.
- Clarity and category fit
- Spelling and pronunciation
- Distinctiveness and memorability
- Channel fit, including voice and word of mouth
- Expansion width beyond the first feature
- Founder resonance, entered by a person rather than inferred
4. Keep hard blockers outside the average
An average can hide the exact fact that should stop a launch. A direct active same-category competitor, a family-of-marks collision, a consistent spoken-name failure, or unfinished legal work must remain visible beside the score. Do not let eight good dimensions mathematically erase one disqualifying risk.
Unknown is also a result. A blocked social platform, missing provider, search ambiguity, rate limit, or timeout should produce a manual action, not an optimistic assumption.
5. Research only the finalists
Deep research is expensive and cognitively heavy. First reduce the field with product fit and calibrated availability evidence. The shortlist itself has no cap; only deep research is bounded, at up to three finalists per run. Investigate those finalists for namesakes, category crowding, linguistic concerns, social access, preliminary trademark signals, and current registrar terms. As of July 2026, a researched finalist costs 250 Naming Credits on the published schedule, which is one more reason to narrow the field first.
The result is still due diligence, not legal clearance. Preserve source links and timestamps so the founder, counsel, or future agent can reproduce the work instead of starting from a mysterious green badge.
6. Export a decision record
A name decision should be reversible when facts change. Export the brief, generated field, shortlist, rejection reasons, evidence, scores, hard blockers, chosen name, and remaining manual actions. This makes later rechecks faster and prevents the team from repeating arguments that were already resolved. When a recheck is due, the domain name checker reruns the same calibrated registry evidence for the chosen name.
The winning word is only part of the asset. Keep the reasoning that explains why it fit the product, what was checked, what was not, and what must happen before launch.