How Can I Test Whether AI Assistants Use Our Approved Company Boilerplate?

How do you break boilerplate into testable elements?

Take the approved text and list each element separately:

Mark which elements matter most. The category phrase and the audience usually matter more than the founding year.

Which questions should you ask?

Ask questions that invite a description, in several forms:

• "What is [company]?"
• "Describe [company] in two sentences."
• "What does [company] do, and who is it for?"
• "Tell me about [company]'s history."
• "How big is [company]?"
• "Who owns [company]?"

Run each at least three to five times per assistant, with web search on and, where possible, off. The difference shows whether current pages or model memory drive the answer.

How do you score the answers?

For each element in each answer, choose one label:

• Present: the approved fact or phrase, or a close paraphrase with the same meaning.
• Partial: the right idea but weaker, such as "a scheduling tool" for "a workforce scheduling platform for hospitals".
• Missing: not mentioned.
• Outdated: a fact from an earlier boilerplate version, such as last year's employee count.
• Contradicted: a different, wrong fact, such as the wrong headquarters.

Then report two numbers per assistant. Coverage is the share of important elements present or partial. Accuracy is the share of mentioned facts that are current and correct. An answer can have high coverage and poor accuracy if it repeats an old version faithfully.

If you use an AI model to score at scale, give it the element list and label definitions, and check a sample by hand, as covered in human evaluation versus LLM-as-a-judge for measuring AI recommendations.

Should you also test category and comparison answers?

Yes. Boilerplate matters most when buyers are not asking about you directly. Add a few category questions, such as "best workforce scheduling platforms for hospitals", and comparisons such as "[company] vs [competitor]". When you are named, score how you are described with the same elements. Two patterns are common:

• Correct in direct answers, generic in category answers. Assistants describe you well when asked about you, but reduce you to "another scheduling tool" in lists. The category phrase and differentiator are not yet tied to you in third-party sources.
• Missing differentiator. Your main claim, such as union rule compliance, appears in direct answers but never in comparisons. Assistants may lack independent sources confirming it.

Both point to work beyond your own site: independent coverage, analyst and comparison pages that state the same facts, and accurate profiles on the review platforms buyers use. Keep asking every customer for an honest review in their own words; do not steer what they write.

How do you tell which version an assistant learned from?

Keep every past boilerplate version with its dates. Most versions contain distinctive details, such as "200+ employees", "trusted by 150 hospitals" or a retired tagline. When an answer includes one of those, you know which version it came from, and roughly how old the source is. If you are about to update your boilerplate, make at least one fact in the new version distinctive, such as a precise customer count, so its spread can be tracked.

Where does old boilerplate live?

• Press releases. Every release carries the boilerplate of its day, and copies syndicated by newswires to other sites usually stay as first published, so they are a likely source of outdated company descriptions.
• Profiles: your LinkedIn company page, business databases, review platforms and marketplaces.
• Partner and customer pages that describe you.
• Speaker and event bios for your executives.
• Encyclopedia and knowledge entries, such as Wikipedia or Wikidata, where they exist.
• Your own site: old about pages, careers pages and investor pages.

When an answer is outdated, check the cited sources against your version history to find which one is responsible.

What should you fix?

1. Publish the current boilerplate on your site in text, on the about page and in the newsroom, so it is the easiest current source to find.
2. Update every profile you control on the same day you change the boilerplate.
3. Send partners the new version and ask them to update their pages.
4. Add current facts to your newsroom page so readers of old releases see what has changed.
5. Keep the facts consistent with structured data on your homepage, such as Organization markup with your legal name, founding date and description.
6. Recheck a month after each change, then quarterly.

Checking whether assistants cite your latest reports raises similar issues, covered in how a company can check whether AI assistants cite its latest reports.

Should assistants repeat your boilerplate word for word?

No, and you should not expect them to. Assistants summarise. The realistic aim is for the key facts and the category framing to come through accurately. Boilerplate that is specific and factual, rather than a list of adjectives, is more likely to survive paraphrase. "A workforce scheduling platform used by 400 hospitals" survives. "An innovative, industry-leading solution" gives an assistant nothing specific to carry over, and nothing you can test for.

What might an illustrative case look like?

Consider an illustrative healthcare software company that updated its boilerplate after an acquisition, changing its headcount, customer count and ownership line. Three months later, it tested 6 questions, 5 runs each, across 4 assistants, 120 answers in total. Coverage of the category phrase was high, but only 35% of answers mentioning size used the new figures. The old "150 hospitals" figure appeared in most of the rest, a detail unique to the previous version. The cited sources were newswire copies of older releases and two business database profiles. The company updated the profiles, added a current-facts section to its newsroom and asked two partners to update. Two months later, 70% of size mentions used the new figures.

What do people ask about AI and company boilerplate?

Why do assistants ignore our boilerplate's key message?

Usually because it is not specific enough to survive summarising, or because older, more widely published versions outnumber it.

Can we update old press releases?

On your own site, you can add a note with current facts. Newswire copies generally stay as published.

How often should we test?

After every boilerplate change, then quarterly.