Generative engine optimization (GEO) agency
The answer is only the surface. We work underneath it.
Underneath is a generative engine optimization agency. We help businesses become easier for AI-powered search and answer systems to understand, retrieve, describe and reference, including ChatGPT, Google AI Overviews and AI Mode, Gemini, Claude, Perplexity and Microsoft Copilot.
When someone asks an AI assistant which company to choose, the answer can look simple. But an answer is built from information: claims, sources, entities, relationships, evidence, pages, reviews, the questions people ask, the information competitors have, and the things that are current, missing or conflicting.
Underneath investigates that information system, changes what matters, and measures what happens next.
Illustrative example · AI answer · 3 sources
- ChatGPT
- Gemini
- Perplexity
- AI Overviews
What’s the best accounting software for a growing business?
For growing businesses, YourBrand is frequently recommended, citing its transparent pricing[1] source 1 and wide range of integrations[2] source 2. Owners on forums also highlight its onboarding support[3] source 3
- yourbrand.com/pricing
- yourbrand.com/integrations
- reddit.com/r/smallbusiness
The AI assistants and AI search we get you mentioned, cited and recommended in
- ChatGPT
- Google AI Overviews
- Google AI Mode
- Gemini
- Claude
- Perplexity
- Microsoft Copilot
- Bing
What happens underneath an AI answer?
A buyer asks:
“What’s the best accounting software for a growing business?”
An AI system may return three recommendations. The customer sees the answer. We look underneath it.
What is the system trying to understand?
Who are the companies? What category are they in? Who are they for?
What is being said?
What claims are being made about each company?
What supports those claims?
Which pages, reviews, publications, communities or other sources provide the evidence?
What is missing or conflicting?
Where is information incomplete, outdated or inconsistent?
Why does one company appear while another does not?
The answer tells us where to look. The information underneath tells us what to change.
Illustrative example. Not a representation of any one platform’s internal process.
Understand first. Then change what matters.
We don’t start with a list of pages to publish. We start with the questions your customers ask and the answers they receive. Then we work backwards.
Diagnose
Find out what the market, and the machines, actually understand about you.
We establish a baseline across AI visibility, search, customer questions, competitors, entities, sources, customer language and technical foundations. We look beyond mentions and rankings to understand the claims, evidence and relationships underneath the answer.
Architect
Decide what needs to become true.
The diagnosis becomes a prioritized plan across content, technical foundations, entities, sources, evidence and measurement. Every recommendation has a reason. Nothing is added simply to create activity.
Deploy
Change the information system around the business.
Depending on what the diagnosis finds, that can mean changing important pages, improving information architecture, strengthening technical foundations, clarifying entities, improving external sources or making important claims easier to support and reference. The work follows the problem, not the other way around.
Compound
Measure. Learn. Change. Test.
AI answers shift. Search results move. Competitors change. Businesses change. So the program does too. Every cycle starts with better evidence than the last.
We don’t optimize one page in isolation.
Your website is one source. Your reviews are another. Your profiles are another. Your customers are another. Publishers, communities, directories and other sources all contribute to the information environment around your business.
We look at how those pieces fit together. Then we identify where the system is helping you, where it is letting you down, and what deserves attention first.
We ask seven questions.
Understanding
Why do your best customers choose you?
Recognition
Do search engines, AI systems and the sources they read recognize who you are?
Retrieval
Can the right information be found when it matters?
Evidence
Are the important claims supported?
Competition
Where do competitors have a stronger information footprint?
Measurement
What actually changed?
Learning
What do we know now that we didn’t know before?
The questions are simple. Answering them well is the work.
We don’t just follow the industry. We study it.
Underneath publishes original research into how AI search and answer systems actually behave. Not opinion pieces dressed up as research. Our studies publish the question, sample, method, limitations and underlying data so the numbers can be checked and reused.
23 studies. One principle.
If we make a claim about how AI systems behave, we want evidence behind it. Our research has examined:
- AI crawlers
- Which crawlers websites block, how sites serve information to AI agents, and how common formats such as
llms.txtand Markdown actually are. - AI search
- How often AI Overviews appear, how their citations relate to traditional rankings, how AI Mode differs from AI Overviews, and which sources appear in answers.
- AI assistants
- How ChatGPT, Gemini, Perplexity and Claude agree, or disagree, on recommendations.
- Recommendation stability
- How answers change when the same question is repeated, reworded or asked in different countries.
- Sources and citations
- Which pages assistants cite, how fresh those sources are, and where those citations come from.
- Local AI search
- How AI recommendations compare with Google Maps and how accurately assistants represent local businesses.
The findings are not a collection of assumptions about what AI “likes.” They are observations from defined datasets.
A few things we’ve learned.
AI recommendations are not fixed rankings.
Across 80 buyer questions, all four assistants we studied recommended the same first pick for only 10.0% of questions.
Rewording the question can change the recommendation.
Adding “on a tight budget” kept the same first brand only 15.3% of the time across three engines in our study, against 68.0% when the same question was simply asked again.
AI citations don’t simply mirror Google rankings.
In our study of 4,051 AI Overview citations, only 28.7% of cited URLs ranked in the top 10 for the same search.
AI answers can use sources outside the traditional top 10.
17.9% of the AI Overviews we studied cited Reddit, and most of the cited threads were outside Google’s top 10 for the same search.
AI answers can contradict a business’s Google profile.
Across 159 local businesses and four AI engines, 18.9% of answers contained a fact that differed from the business’s Google profile.
These are not universal laws. They are measurements from defined studies, with published methods and limitations. That distinction matters.
We measure the work in four layers.
Visibility → Evidence → Behavior → Business. An AI mention is not automatically a business result, so we don’t collapse everything into one “AI visibility score.”
Visibility
What is being said? Mentions, recommendations, descriptions, citations, competitive presence and persistence across important questions.
Evidence
What supports it? Claims, sources, entities, corroboration, conflicts, freshness and information coverage.
Behavior
What did people do? AI referrals, engagement, inquiries, applications, bookings and conversions.
Business
Did it matter? Pipeline, customers, revenue, acquisition cost or the business metric that actually matters to you.
A visibility increase is not automatically a revenue increase. A citation is not automatically a lead. A lead is not automatically a customer. We measure each step on its own terms.
When the work produces results, we show the evidence.
No invented scores. No convenient stories. When client results are available and approved, we show the baseline, intervention, measurement and outcome, not just a screenshot of an AI answer. Until a result is substantiated and approved, we label it accordingly.
Because a good case study should answer four questions:
- What was the problem?
- What changed?
- What happened afterward?
- What can we actually attribute to the work?
Our standard is simple: a result is only proof if you can check it.
Generative engine optimization.
GEO is the work of making a business easier for AI-powered search and answer systems to understand, retrieve, describe and reference. It builds on SEO, but asks a broader question:
Given everything published about this business, on its own site and elsewhere, what will an answer system understand, retrieve and say?
Depending on the diagnosis, our work can include
- AI visibility and answer analysis
- Buyer-question and query research
- Claim and evidence analysis
- Content and information architecture
- Technical GEO
- Entity and information architecture
- Source and evidence strategy
- Monitoring and experimentation
- Team enablement
We don’t prescribe the work before we understand the problem.
Built for businesses where being chosen matters.
Home and local services
Businesses where the recommendation can become a call, booking or visit.
Franchises and multi-location brands
Businesses that need consistent information across markets and locations.
B2B software and technology
Businesses competing for consideration, comparison and shortlist inclusion.
Financial services and insurance
Businesses where accuracy, trust and compliance matter.
Healthcare and dental
Businesses where reputation and local recommendation matter.
Legal and professional services
Businesses where expertise and trust shape the decision.
Retail and ecommerce
Businesses where product information, comparison and recommendation matter.
Hospitality and travel
Businesses where local discovery and recommendation can directly influence demand.
One team. One accountable lead.
Every engagement has one senior lead responsible for the strategy, priorities, decisions and overall program. Specialists join where the work requires them.
You know who is responsible. You know what changed. You know what we learned. And you know what happens next.
Every program runs on the weekly 30, a 30-minute review in the same slot every week, with a decision-maker from your side in the room. If a decision-maker cannot join the weekly review, we may not be the right fit.
Because the visible answer is not the whole problem.
A customer sees a recommendation. We want to understand what produced it.
A system cites a source. We want to understand why that source was there.
A competitor appears in the shortlist. We want to understand what information supports their position.
And when your business is missing, we want to know whether the problem is recognition, retrieval, evidence, representation, or something else entirely.
The answer is only the surface. That’s why we’re called Underneath.
Free strategy call
The market is already asking.
The question is whether your business is part of the answer. Tell us what you’re trying to change. We’ll take a first look at how your business appears across search and AI answers, what information sits underneath that visibility, and where we would start.