AI can help assemble a creator shortlist, but treat every recommendation as an unverified lead until you open the profile and check the evidence behind it. A useful AI influencer discovery workflow records what the account publishes, when you checked it, and which selection criteria remain unknown. A convincing paragraph about a creator does not complete those checks.
Two different workflows sit under the AI label. Modash's discovery-tools article describes searches against creator databases using descriptions or images. A general chatbot can also produce a list of names. Ask which records a tool searched and what each recommendation links to; the label alone tells you little about the evidence.
OpenAI warns that ChatGPT can fabricate references and sound confident when wrong. That supports checking cited material yourself. It does not establish an error rate for any creator-discovery product.
Use the protocol below before a recommendation reaches your outreach list. It is an editorial workflow recommendation, not a platform requirement. The fictional case shows how to handle a plausible creator whose identity cannot be confirmed.
Give the search a testable brief
Start with criteria another reviewer could check. "Find authentic creators who love sustainability" leaves too much room for interpretation. For a hypothetical refillable cleaning product campaign, use a brief like this:
- Content requirement: recent demonstrations of household cleaning routines.
- Audience requirement: evidence of viewers in the market where the product ships.
- Format requirement: examples of explaining a product on video.
- Open questions: availability, rates, and interest in trying the product.
Ask AI to suggest search terms, locate candidates, or summarize material you supply. Require a profile URL and supporting post URLs for every proposed account. Let it return fewer candidates than requested when it lacks evidence. A request for ten names should never become a reason to fill the last rows with guesses.
If the brief produces broad results, use the search-term worksheet to turn the customer's use situation into narrower searches.
Check identity before you score fit
Open the recommended URL yourself. Record the platform, current handle, profile URL, and check date. Confirm that the cited posts belong to that account. A similar display name is insufficient to join two profiles into one person. On YouTube, handles identify channels and differ from channel names. Save the handle and channel link rather than relying on a display name.
When the link fails, mark the candidate unresolved. A failed link alone does not prove that AI invented someone. The account may have changed its handle, removed content, or become inaccessible. Search the platform for the current identity and look for a creator-controlled cross-link before substituting another profile.
Do not silently repair a recommendation by choosing the nearest name match. If you find a different account worth considering, create a new candidate record and evaluate it on its own evidence. It should inherit none of the first record's audience claims, content history, or scores.
This check comes first because evaluating the wrong person makes every later calculation irrelevant.
Break the recommendation into claims
An AI-generated sentence might combine five separate assertions: the creator exists, covers your topic, reaches your customers, accepts partnerships, and can drive sales. Check each assertion separately.
Use three evidence labels:
| Label | Meaning in your review | Example |
|---|---|---|
| Observed | You opened the underlying material and recorded what it shows. | A dated video demonstrates a cleaning routine. |
| Inferred | Someone interpreted evidence or produced an estimate. | A summary classifies the channel as relevant to low-waste households. |
| Unknown | You have no usable supporting evidence. | The creator's availability for next month's campaign. |
A creator statement needs attribution even when you observe it directly. Seeing a location in a bio supports "the bio lists this location." It does not establish where the audience lives. Likewise, seeing a product demonstration supports a content observation; it does not prove buying intent among viewers.
If an analytics provider supplies an audience estimate, keep the provider's name, report date, metric definition, and methodology link beside it. If you cannot establish what it measures, leave that selection criterion unresolved. Do not convert an unexplained percentage into a verified audience fact.
Use platform metric definitions when reading a creator's report. YouTube's Audience help places these reports in YouTube Studio and defines Top geographies by watch time. Do not relabel that as a percentage of followers. YouTube also warns that some demographic and geography data may be limited. If a report omits a field, ask what is available rather than asking AI to fill it.
For a deeper decision about audience evidence, use the audience-fit rubric. It addresses the gap between a creator's subject matter and the people a campaign needs to reach.
Keep a claim-to-source record
Store enough detail for a colleague to repeat the review without asking the AI to recreate its answer. A short record per candidate can use these fields:
Candidate label:
Platform and current profile URL:
Profile checked on:
Campaign criterion:
Exact claim being evaluated:
Supporting profile or post URL:
Post date or reporting period:
What the source supports:
Evidence label: observed / inferred / unknown
Limit or contradiction:
Decision: review / hold / exclude
Reviewer and decision date:Choose a review window that matches the campaign. For a current product launch, this might mean checking recent posts alongside any older examples that explain the creator's style. Record the window and sample size you chose. Do not describe a small sample as a complete content audit.
Separate the date of a post from the date you opened it. For analytics, also record the reporting period. A fresh screenshot of an old period does not make the underlying measurement current.
A fictional recommendation that fails verification
Consider this entirely hypothetical AI output:
Candidate A makes refill-based cleaning videos, has 42,000 followers, and reaches a 78% UK audience. An April product partnership makes them a strong choice.
The account, counts, percentage, and partnership in this example are invented for the exercise. There is no real creator behind Candidate A.
The hypothetical reviewer finds that the supplied profile link does not resolve. Searching the suggested name returns a different account whose posts cover home renovation. The supposed partnership citation opens a brand category page with no identifiable creator post.
The review record should look like this:
| Claim | What the hypothetical review found | Decision |
|---|---|---|
| Candidate identity | No confirmed profile; similar-name result has no established connection | Hold the whole candidate |
| Cleaning videos | No attributable post links | Unknown |
| 42,000 followers | No confirmed profile or dated observation | Remove the count |
| 78% UK audience | No analytics report, source, or methodology | Remove the percentage |
| April partnership | Citation does not support the claimed event | Remove the partnership claim |
In this exercise, the recommendation was fabricated. In a real review, the evidence might establish only that you could not verify it. Record that narrower finding. Do not accuse a real creator of fraud because an AI response supplied a broken link.
Do not average these failures into a low fit score. Without a confirmed identity, there is no candidate to score. Ask for a new, independently supported recommendation and retain the failed record so another reviewer does not revive it.
Decide what can move to outreach
Use these three outcomes consistently:
- Review: identity is confirmed, linked content supports the basic fit, and the record lists remaining questions.
- Hold: a requirement that determines eligibility lacks evidence, or the profile identity remains unresolved.
- Exclude: attributable evidence contradicts a required criterion. Write the specific reason.
Unknown data should stay unknown. It need not disqualify every creator. You may decide that audience evidence can be requested during initial contact, while confirmed identity and relevant content are required before contact. Write that policy before comparing candidates so the same rule applies to everyone.
For candidates that pass, use the defensible-shortlist process to compare them against the campaign brief. Verification establishes what you can rely on; selection still requires judgment about the campaign.
Before sending the next outreach email, pick one AI recommendation and complete its claim-to-source record. If a colleague cannot trace its strongest selling point to evidence, remove that claim or put the candidate on hold.
