Why AI Recommends Some Companies — and Ignores Others

Imagine two Dutch companies that both help SaaS teams improve lead quality. One explains its audience, services and limits plainly. The other says it “transforms growth through innovative solutions.” If you were asked to recommend one to a marketing director with a specific problem, which could you assess? An AI assistant faces a version of that same information problem.

Start with the question, not the chatbot

“What company should I use?” sounds simple, but the useful answer changes with the missing details. An accounting firm suitable for a Dutch freelancer may be wrong for a multinational. A B2B marketing consultancy that works with long sales cycles may be a poor match for a local consumer retailer. Geography, budget, regulation, existing tools and the outcome sought all alter the shortlist.

Different assistants also work differently. Some search the public web during the conversation; some rely on particular search indexes, connected tools or material supplied by the user. Their answers can vary by wording, location and time. There is no single published rule that determines whether ChatGPT, Perplexity, Gemini, Claude and every other system will name a company. OpenAI says ChatGPT search uses multiple factors to seek relevant, reliable information and does not guarantee placement. Google says its AI search features build on Search foundations and likewise do not guarantee inclusion. Those are documented facts; the framework below is our practical interpretation of what a business can improve.

Think of the work as becoming easier to understand → retrieve → verify → compare → cite. Recommendation, when appropriate, is a possible outcome—not the sixth step a marketer can force.

Can someone tell what your company actually does?

Entity clarity is a plain-language issue before it is a technical one. A visitor should be able to identify the company, the specific service or product, the people it serves and the markets it covers. “We help ambitious organisations unlock potential” gives a buyer almost nothing to compare. “Accounting software for Dutch freelancers who invoice in euros and file VAT returns” gives them a starting point. So does “B2B marketing and pipeline measurement for SaaS companies in the Netherlands.”

That does not mean repeating a keyword in every heading. It means choosing a description specific enough that a wrong-fit buyer can rule you out. A provider serving only established B2B SaaS teams should say so. If you advise on strategy but do not run outbound calling, say that too. Clear boundaries are valuable information.

Answer-first pages help the same way. A buyer asking “Why is our Google Search traffic not producing leads?” needs a diagnosis of intent, landing-page fit, conversion and sales handoff—not three paragraphs about your company’s passion. Our guide to search traffic without leads is an example of answering the problem before presenting a service. Good pages can still have a commercial next step; they simply earn it by helping first.

A claim is not the same as evidence

“We are the best” asks the reader to take your word for it. A named methodology, a transparent explanation of what is measured, a real case study with permission, a current certification, a qualified author or a genuine customer review gives the reader something to inspect. Not every company has all of these. A new firm should not invent customers to look established. It can show how it diagnoses a problem, what a deliverable contains, what it cannot promise and which sources informed its advice.

Information outside your own website adds a different kind of context. A relevant association listing, an accurate partner profile, a credible industry directory, a substantive interview or a genuine product review may help a person—or a search system—cross-check who you are. Google’s local-business guidance explicitly describes a mix of business-provided and third-party information and notes the importance of accurate, complete details. That is evidence for Google’s local ecosystem, not proof that every AI model scores all directories or reviews the same way.

The useful question is not “How many backlinks can we buy?” It is “Where would a careful buyer reasonably expect to verify this company?” A software product might need a current product directory and clear documentation. A regulated adviser may need a professional registration. A specialist agency may benefit more from an honest founder profile and well-supported articles than from 50 empty listings. Spam creates the appearance of corroboration without its substance.

Consistency matters here. If one directory says you serve enterprise retailers, your site says early-stage SaaS, and an old profile lists services you stopped offering, a buyer has to guess. The same applies to old prices, addresses, opening hours or product features. Review the few places where people actually encounter you; update them when the business changes. For local companies, keep the Google Business Profile accurate if the business is eligible for one.

Make the information accessible—then describe it accurately

A brilliant page that is blocked, hidden behind a login or absent from the site’s navigation is harder to use. The basics still matter: pages should load, be indexable where intended, connect through sensible internal links, and have a clear preferred URL when near-duplicates exist. A sitemap helps discovery; a canonical signal helps indicate the preferred version. Neither compels an engine to index or cite a page.

For ChatGPT search specifically, OpenAI advises publishers not to block OAI-SearchBot if they want their content available for summaries and snippets. That is different from GPTBot, which concerns potential model training. Check both robots.txt and the actual edge or firewall response: an “allow” directive cannot help if a bot receives a challenge or a 403. Google similarly says pages need to be indexed and eligible for a search snippet to appear as supporting links in AI Overviews or AI Mode. Technical eligibility is the entrance, not the finish line.

Structured data is a label attached to visible information. Organization and LocalBusiness markup can clarify a business identity; Service, Product and Offer may fit appropriate service or product pages; Person can describe a real author; BreadcrumbList can explain a page’s place in the site. Use only what the page actually supports. Google says structured data can help it understand content, but does not guarantee a rich result. It also says there is no special schema markup required for its generative AI search features. So avoid adding invented reviews, fake ratings or a dozen irrelevant types in the hope that a machine will be impressed.

Our agent-readiness audit can expose some of these technical gaps, while the AI growth system shows how automation and measurement connect to broader marketing work. Neither tool can certify that an assistant will recommend a brand. For a wider view, read our AI search visibility guide.

Measure the answer, not just your position on a dashboard

A one-off prompt—“Does ChatGPT recommend us?”—is a poor measurement system. The wording may be unrealistic, the model may answer differently tomorrow, and a founder will naturally ask questions that favour their own brand. Start with the questions actual buyers might ask before they know your name. Include broad discovery, comparison, location-specific and problem-led questions. Leave out brand names unless you explicitly want to test brand understanding.

For example, build a fixed panel of 50 realistic buyer prompts. Record the exact wording, assistant, date, region or language, whether web search was used, the companies mentioned, which ones were linked or cited, and whether the cited source actually supported the statement. If your company appears in 7 of the 50 answers and is linked in 3, those numbers describe your own sampled test. They are not an official ChatGPT rank or a market-wide recommendation rate. The numbers here are an illustration, not Samsina results.

Run the same set again on a schedule, ideally with a consistent method and saved answer snapshots. Keep mentions separate from citations. Also note which competitors appear and which sources the assistant used. A change in the assistant’s model, search index or prompting may move the result even when your website did not change. Small samples should be treated cautiously. Track search impressions and visits alongside this benchmark, then follow the commercial chain: relevant visit, useful action, saved enquiry, qualified lead and confirmed meeting. Do not call an AI mention “pipeline.”

The best question for the team after a test is not “Did our score go up?” It is “What would make the buyer’s decision easier?” If a competitor is cited because its pricing is public and yours is not, decide whether transparent pricing suits your business. If an assistant cannot distinguish your offer from cold-calling agencies, clarify the difference. If a directory carries an obsolete description, update it. Measurement is valuable when it points to a fixable information gap.

A practical way to work: Audit → Fix → Monitor → Prove

Audit. Begin with real buyer questions. Compare the answers with the website and independent sources that are available. Note uncertainty rather than declaring every absence a “technical error.” Check whether the company was a relevant fit for each question in the first place.

Fix. Improve the few things a business controls: an accurate service description, an answer that addresses the buyer’s situation, named and supportable evidence, sensible site links, structured data that matches the page, and access for the crawlers you intend to allow. Seek genuine third-party presence where it helps a buyer verify you. Fix broken customer journeys too; an AI-referred visitor who cannot submit a working form is not a win.

Monitor. Repeat the same prompt panel, watch relevant search queries and AI referrals where they can be identified, and log sources and competitors. Keep a dated record of site changes so you can separate a content edit from unrelated variation in the assistants’ answers.

Prove. Report what actually changed. “We made ten pages easier to understand and fixed crawler access” is a verifiable operational result. “We became the number-one recommendation in AI” would require much stronger evidence and, in most cases, would still be an overclaim. Samsina calls this approach AI Recommendation Readiness because the honest work is to improve readiness, then observe outcomes—not sell guaranteed placement.

If you are curious how your company looks from an AI buyer’s perspective, we can examine the questions customers really ask, show where another source is being used instead, and separate fixable gaps from things nobody controls. A short conversation is enough to decide whether a deeper audit would be useful.

Author: Sina Dadmehr, founder of Samsina. This guide distinguishes platform documentation from Samsina’s working framework; it is not a disclosure of any AI platform’s private ranking system.