Most companies are working on AI visibility backwards. They publish more, they add AI-flavored keywords, and they wait to start showing up in answers.
The waiting rarely ends, because being mentioned and being cited are two different outcomes, and only one of them moves a deal.
Here is the failure mode we see most often. A buyer asks an assistant which tools to consider. Your company gets named in passing. A competitor gets the recommendation, the supporting evidence, and the link. You appear in the answer and lose the answer at the same time. Nothing in a traditional analytics stack will tell you this happened.
This playbook is what we have learned measuring the other side of that answer. Between June 17 and August 25, 2026, Monroya recorded 54,597 citations across 17,103 distinct domains, drawn from 10,282 real buyer questions asked of ChatGPT, Claude, Gemini and Perplexity. The patterns below are the ones that held up across categories, in the order the work is worth doing.
1. Start with the technical foundation
Before you publish anything new, make sure an assistant can retrieve, parse and attribute what you have already built. A page that cannot be fetched is not a weak source. It is not a source at all.
Make sure AI crawlers can actually reach you
Check, in this order:
- robots.txt
- CDN and firewall rules
- Bot protection and rate limiting
- JavaScript rendering
- Your highest-value product and content pages
The trap is assuming that because a page looks right in your browser, a crawler sees the same thing. Fetch it the way a crawler does and read what comes back. A large share of the "we are invisible in AI" problems we investigate turn out to be ordinary retrieval problems wearing a new hat, and they are the cheapest thing on this list to fix.
Lead with the answer, not the positioning
Retrieval lifts passages, not pages. A page that spends three paragraphs on narrative before it answers the question loses to a page that answers in the first forty words, even when the first page is better written.
Put the direct answer at the top. Support it underneath. In our data this single structural change moves more pages into citation range than any amount of schema work.
Implement structured data, honestly
Structured data tells a machine what a page represents. Review whether you are using, and using correctly:
- Organization
- Product / SoftwareApplication
- Article
- FAQPage, where it genuinely applies
Schema does not make a weak page strong. It makes a strong page easier to quote accurately, which matters, because a misquoted page is usually an uncited page. Do not add markup that does not match what a human sees on the page.
Strengthen authorship and expertise signals
Assistants need to know who is speaking and why that source is worth trusting. Make it unmissable:
- Who wrote it
- Who the company is
- What the author actually knows
- Where each claim and statistic came from
- When it was published and last updated
A "Team" byline is not an author. It is the absence of one.
Keep entity facts consistent everywhere
Your company name, category description, founding details and product names should read identically on your site, your profiles and every listing that carries them. Assistants resolve entities across sources. Inconsistent facts produce hedged answers, and hedged answers name fewer brands.
Keep important pages current
Prioritize by value, not by age: product pages, comparison pages, category pages, buyer guides, research, documentation, and the articles already earning traffic. Maintain real publication and update dates, and actually improve the page when you touch it. A stale page that already gets retrieved is worth more than a new page that does not, so refresh before you expand.
What to check
- AI crawlers are not blocked
- CDN and firewall rules are not silently refusing retrieval
- Critical content does not depend entirely on client-side rendering
- Structured data is present and accurate
- Content has identifiable, credentialed authors
- Claims and statistics carry sources
- High-value pages have real freshness signals
2. Stop writing for "AI" and start writing for the buyer's question
There is no AI audience. There are buyer questions, asked in natural language, at four distinguishable moments. We track them as stages because the source mix behind each one is different.
Discovery. "What are the best customer support platforms?" Broad category questions are the hardest ground for a newer company, because assistants already carry strong associations with incumbents. In our window, discovery accounted for 16,237 recorded citations, and it is where challengers are omitted silently.
Evaluation. "Zendesk vs Intercom." "Best alternatives to Salesforce." "Which CRM suits a 50-person SaaS company?" This is the largest stage in our data by a wide margin: 25,501 citations, more than either other main stage. Specificity is where a challenger can win, because the assistant needs structured comparative evidence and there is far less of it in the world than there is category copy.
Decision. "Does it integrate with Salesforce?" "What does it cost?" "What are the drawbacks?" 12,643 citations. Narrow questions with narrow evidence requirements, and the stage where third-party sources carry the most weight.
Retention and switching. "Are there alternatives to X?" "Is X still the best option?" "Should we move off X?" We track this separately from the buyer-journey stages because the asker already has a vendor. It is the smallest measured stage and the most under-served, which is precisely why it is an opening for a challenger. Nobody is writing for the customer who is already unhappy somewhere else.
3. Write the way buyers actually ask
Stop hiding the answer behind the headline.
Instead of:
Transforming the Future of Customer Experience
answer the question:
What is the best customer support software for a 50-person SaaS company?
Then answer it, immediately.
A structure that reliably earns citations:
- The question, stated the way a buyer would say it out loud
- The direct answer, in the first paragraph
- The evidence, explaining why
- The comparison, including the tradeoffs that go against you
- The supporting detail, for the reader who is close to deciding
Also write the variations. "Best X for Y", "X vs Z", and "is X worth it for a team of 20" are three separate retrieval events with three separate source sets. Treating them as one page is how a category-relevant company ends up invisible on the specific question that closes the deal.
4. Comparison content is the most underrated surface in B2B
If you are the challenger, you do not need to win the broadest query first. Start where the decision gets specific:
- You vs. competitor
- Competitor vs. competitor
- Alternatives to competitor
- Best tools for a specific use case
- Best tools for a specific company size
- Feature, pricing and integration comparisons
One condition: the page has to be honest. A thin page asserting that you are better than Competitor X is not a source, it is an ad, and it gets treated like one. Name the cases where the other tool is the right choice. That is the part an assistant can safely quote, and quoting it is what puts your name in the answer.
5. Your own data is the only source type that compounds
Generic advice is infinitely available. Original information is not.
If you have proprietary benchmarks, usage statistics, customer research, surveys, experiments or industry comparisons, publish them. Compare:
Companies should focus on customer retention.
with:
We analyzed 12,000 customer accounts and found that accounts doing X retained at Y.
The first has nothing to lift. The second is a citable object, and other people's citations of it become sources too. Original research is the only asset on this list that other sites, journalists and assistants will carry for you.
6. The four assistants barely agree on anything
This is the finding that surprises teams most, and it is the clearest thing in our data.
Of the 17,103 domains cited across our measurement window, 16,812 of them, 98.3%, were cited by only one of the four assistants. Twenty-three domains, roughly one in a thousand, were cited by all four.
The concentration differs just as sharply. Perplexity drew from 16,344 distinct domains. Claude drew from 65. Same categories, same questions, two completely different retrieval philosophies: one crawls wide and cites the long tail, the other leans hard on a small set of sources it already trusts.
The practical consequences:
- Optimizing for one assistant and assuming the rest follow is the most common wasted quarter in this discipline.
- A single blended "AI visibility score" hides the thing you need to act on.
- Getting into Claude's answers is a source-authority problem. Getting into Perplexity's is closer to a publishing and retrievability problem.
What you actually need to know for any given answer is: which model, which question, which buyer stage, which competitor, which source, and why that source. That is the difference between a dashboard metric and a work order.
7. Look at where the evidence is coming from, because it is probably not you
Across all 54,597 citations we recorded, the source belonged to the company being discussed 0.4% of the time.
Read that again before you approve another blog post. Roughly 99 out of every 100 citations behind an answer about a company came from somewhere the company does not control.
Where they did come from:
- Editorial coverage: 23.1%. The single largest identified category. Third-party writing provides the independent context your own site structurally cannot.
- Analyst and research sources: 13.8%. Durable, slow to earn, hard for a competitor to displace.
- Review and directory platforms: 13.8%. G2, Capterra, TrustRadius and their category equivalents. If your category has an obvious review destination, presence there is a prerequisite, not a growth tactic.
- Social: 7.6% and community: 1.9%. Smaller than the folklore suggests, but unevenly distributed across assistants and disproportionately influential on switching questions. The goal is never to manufacture discussion. It is to be represented accurately where the discussion already happens.
- Documentation and case studies: small in volume, but they answer the narrow decision-stage questions nothing else answers.
Your own domain still matters. It is what makes the rest legible. But the number above should end the assumption that publishing more pages on your own site is the same as building AI visibility.
8. Measure the same questions on a schedule
This is where AI visibility genuinely differs from SEO reporting. There is no stable ranking to check. Models get updated. Retrieval changes. Sources appear and disappear. Competitors publish. The same question, asked twice, can produce two different answers built from two different source sets.
So a one-time audit is close to worthless. Take a fixed set of high-value buyer questions and re-run them on a schedule, asking:
- Did we disappear?
- Did a competitor appear?
- Did the cited source change?
- Did the assistant stop citing our site?
- Did our recommendation position move?
- Did a new source start influencing this answer?
Timing matters when you are measuring the effect of a change. Perplexity re-indexes aggressively and can move within days. Gemini and Claude typically take two to three weeks. ChatGPT is the slowest of the four. If you ship a fix and re-measure on day seven, you measured the wrong week and will conclude, wrongly, that nothing worked.
9. Build the loop, not the checklist
The companies that get this right do not treat AI visibility as a one-time project.
Measure. Identify the questions your buyers actually ask and establish a baseline across all four assistants.
Diagnose. Find where you are recommended, where you are absent, and which competitor is winning each answer.
Trace the source. Determine what is actually feeding the answer. Sort each gap into a technical problem, a content problem or a source-type problem, because those have three different owners and three different timelines.
Act. Fix the technical problems first. They are cheap and they unblock everything after them.
Re-measure. Same questions, three weeks later.
Then run it again. Measure, diagnose, act, re-measure. That loop is the whole discipline.
The takeaway
AI visibility is not about gaming an assistant. It is about becoming the most useful, best-sourced answer to the question your buyer is already asking somewhere you cannot see.
When that buyer asks who they should buy from, it is not enough for the assistant to know you exist. It needs enough evidence to recommend you, and a source to point at when it does.
Want to see where you stand? Run the free check and see which of the four assistants cite you, on which buyer questions, and which sources they are using instead.
Methodology: figures cited above are drawn from Monroya's citation measurement between June 17 and August 25, 2026, covering 54,597 citations from 17,103 distinct domains across 10,282 buyer questions asked of ChatGPT, Claude, Gemini and Perplexity, spanning multiple B2B software categories. Source-type percentages are calculated over all recorded citations, including a 32.5% share our classifier left unclassified, so the identified categories are floors, not ceilings. Results vary by industry, question, assistant and source environment.