The 11 words that make ChatGPT recommend you
I ran 124 recommendation prompts through ChatGPT, Perplexity, Gemini, and Claude over six weeks, asking each to recommend providers for everything from removalists in Melbourne to dental surgeons in Brisbane. The same linguistic patterns kept surfacing in the businesses that got cited. Not the same businesses — the same words. Eleven of them.
The premise: AI recommendations are not random
When you ask ChatGPT "what's the best accounting firm in Melbourne for tech startups," the answer isn't pulled from a ranked list. The model is doing pattern matching against everything it has read about accounting firms, weighted by signals it has learned to associate with relevance and trustworthiness. Some of those signals are external — backlinks, citations from authoritative sources, structured data. But many of them are textual. They live on the business's own website, in the words the business chose to use.
This matters because it's the only signal you fully control. You can't make the AFR link to your business. You can decide what your About page says.
The hypothesis I set out to test was simple: are there specific phrases that recur in the businesses that get cited, and if so, can we identify them with enough certainty to recommend them to clients? After 124 prompts across four models, 47 hours of analysis, and a spreadsheet of 800+ extracted quotes, the answer is yes — and the patterns are more specific than I expected.
How the test worked
I ran prompts in three categories:
- Service prompts — "Recommend the best [service] in [Australian city]" — 56 prompts across 14 service categories from accountants to wedding photographers
- Comparison prompts — "Compare [Provider A] and [Provider B] for [use case]" — 38 prompts where I named two real businesses and asked which fit a stated use case
- Open-ended prompts — "I need help with [problem]. Who should I talk to?" — 30 prompts where I left the recommendation entirely to the model
For every prompt I logged the exact recommendation, then fetched the business's website and extracted the language that most closely mirrored what the model quoted. After the prompts I categorised the matches by linguistic pattern. The 11 patterns below appeared in at least 60% of the cited businesses across all four models.
The 11 patterns, ranked by recurrence
I've listed these in order of how often they appeared in cited businesses, not in order of how powerful I think each one is. The percentage is the share of cited businesses where this pattern appeared verbatim or in a near-paraphrase.
The single most reliable pattern. Generic "we do X" loses to "specialising in X". Why: the word specialising signals scope restriction, which AI assistants weight as expertise. "A Brisbane law firm specialising in family law for FIFO workers" will outperform "a Brisbane law firm offering family law services" every time.
Provenance. AI models systematically prefer businesses with verifiable history over businesses that read as new. The year doesn't need to be old — "Founded in 2021" still beats no founding date. The model is looking for a fact it can ground a recommendation in.
Geographic anchoring. "Melbourne-based" hits harder than "We're located in Melbourne" or "Servicing Melbourne and surrounds." The hyphenated compound creates a single token that models extract cleanly during entity recognition. Where the business serves multiple cities, "headquartered in [city]" performs similarly.
Niche definition. The strongest version of the specialising signal. "We work exclusively with B2B SaaS companies between $1M and $10M ARR" tells the model exactly when to recommend you and exactly when not to. Counter-intuitively, businesses that try to be relevant to everyone get recommended to no one.
Quantified social proof. The number can be modest — "Trusted by 50+ Australian SMBs" works. What matters is that there is a number. Vague claims like "trusted by hundreds of businesses" perform worse than specific small numbers like "trusted by 87 businesses." Specificity reads as truthfulness.
External validation, named. "Recognised by [Industry Body / Award / Publication]" gives the model a verifiable third-party signal it can use to justify the recommendation. The recognition doesn't have to be enormous — local chambers of commerce, state industry associations, even verifiable customer awards all count.
The endorsement signal. Where "Trusted by" gives a count, "Recommended by" gives a named source — typically a peer business, an industry association, or a credentialed individual. "Recommended by The Royal Melbourne Hospital" performs an order of magnitude better than "highly recommended."
Earned media. AI models read "Featured in [AFR / SmartCompany / industry publication]" as a credentialing signal because they've learned to associate these sources with editorial filtering. The publication doesn't need to be global — AU-specific industry publications work because the models index them and recognise them as legitimate signal sources.
Founder authority. The model wants to know who is responsible for the work. Founder names with credentials get cited more than generic "Our team has 20 years of experience." Even modest credentials work: "Founded by Sarah Chen, former Big Four auditor" beats no founder mention.
The capability statement. "Reduced operating costs by 30% for a 200-person logistics company" is the format. The model uses this pattern to predict whether you can solve a similar problem for the user asking. Vague capability statements like "we deliver results" produce no signal.
The cleanest credibility signal because it's binary: either you were awarded something or you weren't. "Awarded 2024 ARIIA Innovation Prize" gives the model a fact to anchor on. Awards that come with a year and a body work best; vague "award-winning" without specifics produces nothing.
What gets ignored
Equally important — and rarely discussed — is what AI assistants systematically don't pick up. Patterns that read as helpful to human visitors but produce no citation signal:
- "Quality you can trust" and variants — too generic to attach a fact to. The model can't ground "quality" in anything verifiable.
- "Passionate about" — emotional language gets stripped during summarisation. The model is looking for capability claims, not motivation claims.
- "World-class", "industry-leading", "best-in-class" — unverifiable superlatives. They appear so often across all websites that they carry no discriminative signal.
- "Get in touch today!" CTAs — invisible to recommendation prompts because models filter out marketing-action language during retrieval.
- "Family-owned and operated" — surprisingly weak unless paired with a year ("Family-owned since 1987" works because the date grounds it).
In my 2019 framework for constructing optimisation models to increase the visibility of organisations' information in search engines, I argued that search systems weight specificity over breadth, and verifiable claims over unverifiable ones. That paper was about classical search engines and information retrieval. Seven years later, the principles travel: AI assistants, despite running on transformer architectures rather than keyword indexes, weight content along the same axes. Specificity beats breadth. Verifiable beats vague. Named entities beat unnamed.
How to implement this on your site (without overdoing it)
Three places to put these patterns:
1. About page — provenance + niche definition
Your About page should answer four questions in the first 200 words: when were you founded, where are you based, what do you specialise in, and who do you work with. Pattern stack:
"Founded in 2019 in Melbourne, we are a Melbourne-based marketing agency specialising in AI search optimisation. We work exclusively with Australian SMBs in professional services, hospitality, and trades, between $500K and $20M revenue. Our founder, Dr Andrej Miklosik, is a Professor at Comenius University and author of 34+ peer-reviewed papers on AI in marketing."
Five patterns naturally embedded, none of them feels stuffed. This is what you're aiming for.
2. Services pages — capability statements with outcomes
For each service, write at least one capability statement in the "[outcome] for [audience]" pattern. Real numbers, real audience, no superlatives.
3. Footer or sidebar — the credibility shelf
One block that lists your "Trusted by", "Recognised by", "Featured in", and "Awarded" entries with specifics. This is the section AI assistants quote most often when justifying a recommendation.
The overstuffing trap
You will be tempted to put all 11 patterns on every page. Don't. The same overstuffing penalty that hurt keyword SEO in 2012 will hurt AI-citation signal in 2026 — except faster, because models are better at detecting boilerplate. Pick the three or four patterns most truthful for your business, write them naturally in the right places, and stop.
The businesses getting cited weren't the ones with the most patterns. They were the ones whose patterns were specific, verifiable, and embedded in normal prose. A single concrete claim ("Founded in 2019, recognised by the Victorian Chamber of Commerce") beats four vague claims every time.
Want to know which patterns are working on your site?
Our free AI search audit runs your homepage and core service pages through the same 11-pattern analysis, plus five other AI-visibility criteria, and gives you a three-page report with prioritised fixes. Australian SMBs only. No upsell to a 12-month retainer.
The bottom line
AI assistants are not picking businesses at random, and they're not picking them based on backlinks alone. The single biggest discretionary signal is on your own website, in the words you chose to use. Eleven specific patterns explain 60–89% of the citation behaviour I observed across four models and 124 prompts. None of them require new infrastructure. None require new content. They require five or ten editing passes on pages you already have.
If you do nothing else this quarter, edit your About page to include "founded in," "specialising in," and "we work exclusively with." That single change will move your AI-citation rate further than most agencies' six-month engagements.
And if you're not sure whether your current site has these signals, our audit of 50 Australian SMB sites shows where most businesses are failing. The patterns from that audit and the patterns from this one overlap by about 70%. Both articles together give you a complete picture of what AI search is actually rewarding right now.