The questions in an AI visibility report can decide which pages a team rewrites, which publishers it approaches and which product claims it tries to strengthen.
If those questions come from an unsupported idea of what buyers care about, that assumption can become an entire strategy.
The same questions then become the measure of success. The team improves its answers, the score rises and the work looks justified. Yet nobody has checked whether the questions reflect the decisions its customers need to make.
That is my main concern with poorly researched prompt sets. They can send a capable team into months of work on the wrong problem.
A small comparison I ran in web hosting illustrates the risk. Questions built from customer reviews often specified what a good service would need to do. Questions generated directly from the category more often asked for a general judgement: the best value, the most reliable host, the best support.
Those differences suggested quite different priorities for the business.
The prompts become the plan
Ask which hosting provider has the best support and you might get a list based on reviews, rankings and reputation. If your brand is missing, the obvious response is to improve its presence in those sources.
That may be useful work. But suppose your prospective customers are worried about whether a support agent will actually resolve a technical problem. A programme built around broad support rankings may give that concern very little attention.
The choice of prompt has already influenced the diagnosis and the proposed work.
I would expect a team to justify that choice before committing its budget. Why this question? Whose concern does it represent? What evidence suggests that answering it will help the people we want to reach?
A plausible question is a reasonable starting hypothesis. It needs more support before it determines the strategy.
“Good value” can mean knowing what the plan actually includes
A Hostinger reviewer wrote: “To unlock more flexibility, it seems like you need to move to higher-tier plans”. A SiteGround reviewer described tools that “were once included with hosting” but were “now locked behind pricing tiers”.
Those comments helped produce this research prompt:
Which entry-level hosting plans don’t lock basic features behind upgrades?
The nearest question in the directly generated set was:
Which web hosting companies offer the best value for money?
Hostinger was recommended on the broad value question by all three engines in the test.
The entry-plan question produced a more specific assessment. ChatGPT built a table of plans, inclusions and renewal prices. It recommended SiteGround StartUp, GreenGeeks Lite and KnownHost Basic. Its answer cited Hostinger’s own site when describing limits on included email and backups.
Perplexity kept Hostinger Premium as an option but qualified the recommendation with its weekly backup provision. Google still recommended Hostinger alongside DreamHost.
For a strategy, I would take a different next step from those answers. I would check which features the intended customer considers essential, what each plan includes and whether the engines have represented it correctly.
If the information is unclear or misread, that gives us a content problem to address. If the entry plan genuinely lacks something those customers need, the decision concerns the offer and its suitability for them.
A stronger “great value” message would not settle that question. Neither would a favourable score on broad value-for-money prompts.
“Good support” can mean someone who will fix the problem
A WP Engine reviewer praised access to someone who could “actually solve your problem, not just read from a script”. Other reviews described standard responses and first-line agents who could not resolve an issue.
The research prompt asked:
Which web hosts have support agents who actually fix technical problems instead of reading from a script?
The directly generated question asked:
Which web hosting companies have the best customer support?
On the specific question, ChatGPT led with KnownHost and also named Rocket.net, BigScoots and Kinsta. Its sources included providers’ support-scope pages. Perplexity used customer reviews to recommend InMotion, SiteGround and Krystal.
On the broad question, ChatGPT led with SiteGround. Perplexity led with IONOS, citing a CNET support rating.
The distinction matters when deciding what to improve. A high support rating gives a general endorsement. A support-scope page can explain whether the agent will investigate the customer’s website, which problems they will work on and where their responsibility ends.
The work I would investigate here includes how clearly that scope is explained, what customers say about resolution and whether the service meets the expectation being set.
That might produce a different content brief. It might also identify something that marketing needs to discuss with the support team. A visibility programme should be capable of reaching that conclusion.
“Reliable” can mean warning me before switching my site off
A Namecheap reviewer wrote that “they abruptly turned it off without warning”. A HostGator reviewer described a deleted site following a billing failure and said: “Those emails never came to us.”
The research produced:
Which web hosts don’t suspend or take down sites without warning?
The directly generated question was:
What is the most reliable web hosting service for a business website?
The broad question produced answers about uptime. ChatGPT led with Liquid Web and its advertised uptime guarantee; Perplexity led with GoDaddy, referring to a ZDNET uptime test.
The suspension question produced answers about notice provisions and exceptions in providers’ published terms. ChatGPT began by saying that no host guarantees it will never suspend or remove a site without warning.
The practical concern was whether the customer would be warned and given a chance to act. Uptime statistics do not answer it.
A strategy based on the broad question could lead to more work promoting uptime and general reliability. The customer-derived question gives the team a reason to examine notification processes, suspension policies and how those are explained.
Both sit under “reliability”. They call for different work.
Some relevant questions will never produce a list of providers
One Hostinger reviewer described slowdowns when another site on a shared server received a traffic spike.
The resulting prompt asked:
Why is my site slow on shared hosting when other sites on the server get traffic?
All three engines explained resource contention, often called the “noisy neighbour” effect. None recommended a host. Their sources included Microsoft Azure documentation, Namecheap’s resource-limit guide and StableHost’s knowledge base.
The directly generated set contained no equivalent troubleshooting question.
If a team tracks only questions asking which provider to choose, it will miss this kind of answer and the documentation behind it. A useful technical explanation might deserve attention even though the immediate result is not a recommendation.
Namecheap was visible as a source here. That is different from being recommended as the product to buy, and a report should distinguish the two.
It also requires judgement. I would want to know whether the question reaches an audience relevant to the business and what useful contribution the company can make. An answer without a recommendation is not automatically a commercial opportunity.
The language reveals what the research has retained
In the 20-prompt comparison, 15 directly generated prompts contained “best”, “most reliable”, “fastest” or “easiest”. Only two research-derived prompts used those expressions.
The difference remained within questions asking for a host. Of twelve research-derived questions explicitly asking which provider or plan to choose, ten specified a feature, commercial condition or service behaviour.
Often the important words were the conditions: “without manual setup”, “before I notice them”, or “don’t lock basic features behind upgrades”.
Those phrases retain something a customer wants to avoid or expects the service to do. They give a team a more precise question to investigate than whether the company is generally considered good.
There were overlaps. Both sets asked almost the same value-for-money question, and the directly generated set included concerns about renewal fees and automatic backups. The point is to examine what the set covers and why those concerns were selected.
The reporting can improve while the original assumption goes unchecked
The score differences illustrate how much depends on the selected questions.
| Brand | Recommended in research-derived answers | Recommended in directly generated answers |
|---|---|---|
| Hostinger | 47% (27 of 58) | 57% (34 of 60) |
| SiteGround | 29% (17 of 58) | 55% (33 of 60) |
Here, “recommended” includes being presented as a possible option, rather than mentioned in passing. Two research-derived answers could not be scored for recommendations.
Hostinger’s score was higher on the directly generated questions, while SiteGround’s recommendation rate nearly doubled and came close to Hostinger’s.
These were different prompt sets, not a measure of either company’s progress. But they show why I would want to inspect the questions before deciding what a visibility number means.
The strategic risk goes further. A team can select questions, optimise for them and report improved performance against them without revisiting the assumptions behind the selection.
That can leave a business with a convincing report and an insufficiently researched plan. The amount of effort put into executing the work does not resolve that weakness.
Treat prompt selection as part of the strategy
Before I would approve a programme of AI search work, I would want to see the buyer evidence behind its main priorities.
Reviews are one source. Sales conversations, interviews, support questions, community discussions and search data can contribute too. The task is to identify concerns worth addressing and decide which matter to the intended audience.
For a few important prompts, I would follow the full sequence: what the customer described, how it became a question, what the engines answered and what work the team proposes as a result.
That last step deserves particular scrutiny. Does the proposed work answer the concern in the source? Will another comparison page help, or does the answer point to an unclear plan, a poorly explained service or an expectation the product does not meet?
The sources cited by an engine need checking too. A provider’s own documentation can be useful, as can independent reporting and customer reviews. The presence of a citation does not establish that the answer is right.
Once the work starts, keep a consistent set of prompts for reporting and record changes. Review visibility alongside business evidence such as enquiries, sales conversations and self-reported discovery.
I would rather find a weakness in the prompt research before commissioning the work than explain a rising score that nobody can connect to a useful result.
Ask where the questions came from before agreeing what to do about the answers.
A short note on the comparison
The research used 290 G2 reviews and search demand to build 85 prompts. Twenty were selected at random before reading ChatGPT’s directly generated set of twenty. AI helped write both sets; the research supplied one with customer evidence.
Each ran on ChatGPT, Perplexity and Google AI Overviews, once per prompt per engine. The research-derived answers were collected on 5 October 2026 and the directly generated answers on 15 October. The reliability and support examples above come from the wider research, outside the twenty-prompt sample.
This is a small illustration of the strategic issue, with single answers collected on different dates. It does not establish how often the same differences would recur.
If you would like a second opinion on the questions behind your AI search strategy, get in touch.