Being found was SEO.Being named is a different job.
AIRO is our term for the work of getting named inside an answer rather than listed among the links beneath it, and it is published in full as a free public rubric rather than described. The B2B and eCommerce problems are different, but both start with the same question: can the system identify, understand and trust what you are?
The catalog is the asset.
A shopper asks an assistant what to buy before reaching a product page. Whether your products are named depends on a structured layer most catalogs do not expose: collection, material, fit, occasion, use case, and the relationships between products. The work is building that layer and owning it.
The shortlist forms first.
Ninety-four percent of B2B buyers used generative AI in their last purchase, and ninety-five percent of the time the winner was already on the Day One shortlist. That list is assembled inside an assistant, before a vendor knows it is in play, out of whatever the engines can cite about you.
How an answer engineactually decides who to name.
This is the part most firms keep vague, and it is the part a technical evaluator needs before anything else in this section is worth reading. Nothing below is proprietary. It is how the systems work, and knowing it is enough to audit your own site or to interrogate whoever you hire.
Retrieval, not memory
The model is not recalling your brand from training. On any current assistant a commercial question triggers a search, and the model composes an answer from the handful of documents that come back — typically two to seven sources. So the fight is not for a place in the weights. It is for being one of the documents retrieved, and then for being the easiest one to quote.
Resolution decides whether you are eligible
Before a system can cite you it has to be sure who you are. It reconciles your name against an entity it already holds, using your Organization node, your sameAs edges, and third-party corroboration — directories, review platforms, press. Two companies with similar names and no distinguishing structured data collapse into one, and the one that gets described is whichever had more corroboration.
Extractability beats eloquence
Given two adequate sources, the one that gets quoted is the one whose answer sits in a single self-contained paragraph near a matching heading. A brilliant argument spread over four scrolling sections loses to a plain forty-word answer under a question-shaped h2. This is why the FAQ blocks on this site are written to stand alone out of context, and why they are generated from one file rather than typed twice.
What you can actually measure
Citation share against a named competitor set, on a fixed query list, sampled repeatedly — because these systems are non-deterministic and a single run tells you nothing. Then AI-referred sessions in analytics, which most brands are not segmenting yet. Anything else on offer in this category is a proxy for one of those two.
What this cannot do: make a model prefer you on the merits. If three competitors are genuinely better documented, better reviewed and better known, structure will not invent a reason to name you. It removes the case where you lose for a reason that has nothing to do with the product.
Five layers,and the order is not optional.
This is the whole architecture rather than a menu to choose from. You cannot fix an entity you have not measured, and a catalog marked up underneath a brand the engines cannot resolve is just invisibility with better plumbing. Every line below carries its real status, including the ones we do not sell yet.
Measure
Establish what the engines currently say, before touching anything. Every later decision is scored against this baseline.
- AI visibility baseline across ChatGPT, Claude, Perplexity, Gemini and Google AI OverviewsRunning
- Buyer query map spanning product, brand, category and intentRunning
- Citation, mention and miss log, scored by query type and competitive contextRunning
- Competitive benchmark against three to five brands in adjacent territoryRunning
- Schema and entity audit of the current siteRunning
- Share-of-answer tracking against a named competitive set, month over monthBuilding
- Attribution for AI-referred sessions and the revenue behind themTo add
Entity
Make the brand a thing the engines can resolve. An engine will not recommend what it cannot identify with confidence.
- Founder and executive knowledge panel claim and optimizationRunning
- Canonical bios, consistent across every property the engines readRunning
- Wikipedia eligibility assessment and submissionRunning
- Organization schema deployed site-wideRunning
- llms.txt at the domain root, stating what the brand is and how to cite itRunning
- Citation authority program: press, industry publications, review aggregators, business profile depthBuilding
- Entity reconciliation across Wikidata and the major knowledge graphsTo add
Structure
The engineering layer. Most of the volume in any AI foundation sits here, and it is what makes everything above it compound.
- Product schema across one hundred percent of the active catalogRunning
- FAQPage schema on priority product and brand pages, written in the brand voiceRunning
- Answer-shaped product and listing copy, phrased the way buyers actually askRunning
- Hub page architecture for the comparison queries engines answer fromRunning
- Platform engineering: Shopify theme and metafields, HubSpot CMS, feature branch to staging to productionRunning
- An SEO/AEO writing guide so the client's own team keeps producing citable pagesRunning
- QA across all five engines, plus Rich Results and Schema.org validationRunning
- Prerendered, crawlable HTML so engines that do not execute JavaScript still read the pageTo add
- Agent-readable commerce feeds and an MCP endpoint over the catalogTo add
Systems
AI inside the product, not only in front of it. The claim that a company is AI-augmented is only credible if the service behaves that way.
- Custom assistants built on retrieval over a first-party corpus, not a prompt over a general modelRunning
- Retrieval-augmented search that answers a sentence rather than a filter setRunning
- Orchestration and observability: model routing, caching, cost and latency measured per callRunning
- Brand-voice authoring consoles with a human review gate before anything publishesRunning
- Evaluation harnesses and regression tests, so a model change is caught before a customer finds itBuilding
- AI-augmented operations inside the client's own service: generated reports, drafted communications, predictive flagsTo add
Operate
A foundation nobody maintains is a foundation that decays. This is the layer that decides whether any of the above is still true next year.
- Monthly citation-share read across all five enginesRunning
- A twelve-month AIRO roadmap with quarterly priorities and decision pointsRunning
- Handoff documentation and recorded walkthroughs covering schema, content systems and identity sourcesRunning
- Three-month post-launch read: citation share, AI-referred sessions and revenue against the baselineRunning
- Continuous authority and content production against the roadmapBuilding
- Alerting when a model update moves the brand's position in a tracked answerTo add
Five systems,not five claims.
Each one is deployed, has a client’s name on it, and is written up somewhere on this site. Read the architecture rather than taking our sentence for it.
The orchestration layer underneath the rest of this page. Routing across frontier models, retrieval over first-party data, caching, and observability, with cost and latency measured per call.
Read the architecture ⟶A generative assistant for the wine industry, built on a retrieval layer over a domain corpus rather than a general model with a prompt. Brand, interface, and platform in one engagement.
See the engagement ⟶A live agent that prices a one-of-one instrument from its own inventory history and market signal, and shows the comparable instruments it reasoned from so a specialist can overrule it and the override is recorded.
Read the case study ⟶An internal appraisal tool that prices inventory against comparable sales, rebuilt for speed and accuracy with an AI orchestration layer underneath, and migrated to new infrastructure with no interruption to the team using it daily.
Read the case study ⟶An AI sales assistant that combines several tools into one outreach surface, so a rep sends personalized communication without leaving the record they are working in.
See the engagement ⟶A live calculator that turns a prospect's own numbers into a defensible return figure and writes the result into HubSpot, so the conversation a rep picks up already contains the case the buyer built for themselves.
See the engagement ⟶Merchant quoting rebuilt and running: prices held as data a system can query rather than text somebody has to read, merchant records that carry their own quote and support history, an approval path the business can reconfigure without a developer, and reporting across all of it. The layer that reads that history to recommend a price and score a quote before it is sent is phase two.
See the engagement ⟶A premium home-watch service on Kiawah, for owners who would rather not choose between checking in constantly and not checking in at all. Vellum is where an owner and an operator see visits, tasks and property status as they happen; The Register is the recorded history of those visits, built as its own component and surfaced publicly. Both were prototyped end to end first, so the structure was settled with stakeholders before engineering was committed.
See the engagement ⟶Customs and import compliance, where the question is always the same and always urgent: where is this filing. Two portals on one structure — shippers watch a filing move from submission to release, and operations staff watch that same lifecycle across dozens of accounts at once, filtered by account and searchable by BOL, ISF or Entry number. An AI assistant is being built into the operations view so staff can work across accounts by asking rather than by navigating.
See the engagement ⟶Four positionswe will not move off.
These four decide whether an AI system is still working in eighteen months. They are also the four questions worth putting to any firm you are evaluating, including this one.
Own the layer, do not rent it
The subscription tools run on the same frontier models everyone else uses, and your data sits on their servers. We build the layer on your infrastructure and draw every model through its API. If a model is deprecated, the router moves. If you fire us, you keep all of it and it keeps working.
A human signs off before anything publishes
Generated content passes a review gate owned by a named person on your team. Nothing reaches a storefront, an inbox or a buyer without somebody approving it. That gate is built into the product rather than promised in a policy document, because the failure everyone is worried about is real.
A principal at the front and the back
The person who scoped the system is accountable for what it does in production. AI work fails most often at the seam between the demo and the deployment, which is exactly where a named owner stops being a formality.
Measured like anything else we ship
Cost per call, latency, citation rate, and the revenue or pipeline it is actually judged on. If a model gets cheaper on Tuesday the router moves, and the same numbers are still there on Wednesday to prove it did not get worse.
Questions we getbefore the first call.
What is answer engine optimization?
Answer engine optimization is the work of getting your company named inside an AI-generated answer rather than listed among the links beneath one. It covers four things: measuring where you currently appear across the major assistants, making your brand a resolvable entity those systems trust, structuring your site and catalog so they can be read and cited, and monitoring citation share over time.
How is AEO different from SEO?
SEO competes for position on a results page with ten links. An answer engine cites two to seven sources in a single response and the shopper or buyer often never sees a list at all. Ranking eleventh used to be disappointing; being the eighth source an assistant would have cited is invisible. The technical foundations overlap, but structured data, entity clarity and answer-shaped content matter far more than they did.
How do we find out whether AI assistants are recommending our brand?
We run real buyer queries from your category through ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews, log every citation, mention and miss, and score the results against three to five competitors you name. You get a baseline report with screenshots and direct citations, plus an audit of what is missing from your site's structured data. That is where every engagement starts.
How much does AI visibility work cost?
A full foundation — audit, brand entity, schema across the catalog, and a twelve-month roadmap — typically runs $25,000 to $50,000 depending on catalog size and platform. A standalone visibility baseline is a smaller engagement and a reasonable place to start if you want the measurement before the build.
How long before we see results?
The foundation itself takes six to eight weeks across four phases. Citation improvements begin appearing within that window on queries where the gap was structural. We schedule a formal read at three months post-launch measuring citation share, AI-referred sessions and revenue against your baseline, because that is the honest interval for judging it.
Can you build a custom AI assistant on our own data?
Yes. Seven are in production with clients’ names on them, and two more are earlier and labelled as such. They include an AI sommelier for Aivin, a valuation and pricing agent for Carter Vintage Guitars, an ROI calculator wired into HubSpot for Vodori, a sales assistant for Relat, and retrieval search across a one-of-one guitar inventory for The North American Guitar. They are built as retrieval over your data rather than a prompt on top of a general model, which is what makes the answers specific to your business.
Will we own the AI system you build?
Yes, entirely. It runs on your infrastructure and draws frontier models through their APIs, so the application layer, your data and the retrieval corpus stay with you. If a model is deprecated, the router moves. If our relationship ends, you keep all of it and it keeps working.
Does AI write our content without anyone checking it?
No. Anything generated passes a review gate owned by a named person on your team before it reaches a storefront, an inbox or a buyer. That gate is built into the product rather than promised in a policy document, because the failure mode everyone is worried about is real.