What Does Refusal Mean in AI Search, and How Is It Different From a No-Recommendation?
In AI search, a refusal is when an assistant declines to answer, or declines to name specific providers, usually because of safety rules or caution in areas such as health, law, finance or personal safety. A no-recommendation is different: the assistant does answer, but names no brand. It may explain what to look for, give general advice, or say it lacks reliable information about local providers. A third outcome is a clarifying question, such as "Which city are you in?". None of these is the same as an answer that recommends brands and leaves you out. Code them separately. If you count refusals and no-recommendations as omissions, your visibility figures will look worse than they are and point you at the wrong problems.
How do you tell the outcomes apart?
| Outcome | What it looks like | Example |
|---|---|---|
| Recommendation | Names one or more brands or providers as options or picks | "Good options include [A], [B] and [C]." |
| No-recommendation: criteria | Explains how to choose, with no names | "Look for a licensed installer with at least five years' experience and written warranties." |
| No-recommendation: insufficient information | Says it cannot find reliable providers | "I don't have reliable information about installers in [town]; check local reviews." |
| No-recommendation: redirect | Points to a directory, map or professional body | "Search your state's licence lookup or a local directory." |
| Clarifying question | Asks for more detail before answering | "What's your budget, and which area are you in?" |
| Refusal | Declines to answer or to name providers | "I can't recommend a specific lawyer, but here's how to find one." |
Some answers combine outcomes, such as criteria plus a refusal to name names. Code the main outcome, and note the secondary one.
Why do assistants refuse or avoid recommending?
- Safety and professional caution. In health, legal, financial and security topics, an assistant may decline to give personalised advice and suggest licensed help instead.
- Thin information. For small towns, niche trades or new categories, assistants may not find enough reliable sources to name anyone. That produces "insufficient information" answers.
- Ambiguity. Questions without a location, budget or need invite clarifying questions.
- Phrasing. Wording can matter. In your own runs, compare "Who is the best [professional]?" with "Which [professionals] in [city] offer [service]?", since superlative questions about licensed professionals are the kind assistants are more cautious with. How much wording changes answers is covered in whether buyer phrasing changes AI recommendations more than the platform.
- Model updates. A new model or policy change can shift a whole category from recommending to declining, or back.
How does this change the numbers?
Consider an illustrative law firm tracking 40 prompts, five runs each, 200 answers:
| Outcome | Answers |
|---|---|
| Recommendation, naming at least one firm | 110 |
| No-recommendation (criteria, insufficient information, redirect) | 55 |
| Clarifying question | 20 |
| Refusal | 15 |
The firm is named in 40 answers. Its appearance rate across all 200 answers is 20%. Across the 110 answers that recommend any firm, it is 36%. Both numbers are useful, but they answer different questions. The first says how often a buyer asking these questions would see the firm. The second says how the firm competes when assistants do name firms. Counting all 160 answers without the firm as omissions would suggest an 80% omission rate. The conditional omission rate, among recommendation answers only, is 64%, as covered in what an omission rate is and how it is calculated.
Report the outcome mix itself too. If refusals rise from 8% to 25% after a model update, every brand's visibility falls, and the right response is different from losing ground to a competitor.
What does each outcome mean for strategy?
- Recommendations that exclude you: a competitive gap. Look at the cited sources and competitors, and work on content and third-party evidence.
- Criteria answers: an opening. Make sure your pages show that you meet the criteria assistants list, such as licences, experience, guarantees and prices. Criteria answers today can become recommendations when the assistant has better evidence.
- Insufficient-information answers: a data gap for the whole area or category. Complete profiles, directory listings and clear local pages help, as covered in which missing facts stop AI assistants recommending a business.
- Redirect answers: the assistant trusts a directory or register more than itself. Make sure you are listed there, correctly.
- Clarifying questions: test follow-up answers too. What an assistant says after the buyer gives a location or budget is where the recommendation happens.
- Refusals: no brand wins directly. Focus on being cited for factual and informational questions in the category, and on the directories and registers assistants point to.
How should you code answers consistently?
- Write short definitions for each outcome with two examples each.
- Label from the answer text, not from the sources cited.
- Have two people label a sample and compare, then refine the definitions where they disagree.
- If you use an AI model to label, give it the definitions and examples, and check a sample by hand every cycle.
- Record secondary outcomes, such as criteria plus refusal, so mixed answers are not lost.
What might an illustrative case look like?
Consider an illustrative financial planning firm whose AI visibility "fell by half" in one month. Coding the answers showed the firm's share of recommendation answers was unchanged at about 30%. What had changed was the outcome mix. After an assistant update, refusals on "best financial adviser in [city]" questions rose from 5% to 40%, with the assistant pointing users to regulator registers instead. The firm checked its entries on the registers the assistant pointed to, found an outdated address and corrected it. It also added pages answering the criteria assistants listed, such as fee structure and qualifications. Its reports now show the outcome mix beside visibility.
What do people ask about refusals in AI search?
Can a business do anything about refusals?
Not directly, since they come from the assistant's policies. Make sure you appear correctly in the registers, directories and factual sources assistants point people to instead.
Are clarifying questions a good sign?
They are neutral. Test the follow-up: give the location or need the assistant asked for, and record whether you are then recommended.
Should refusals be excluded from visibility reports?
Keep them in the outcome mix, but calculate competitive measures only on answers that recommend brands.
Published by The Kovmere Team · updated Sep 29, 2026 · More resources