An AI search optimization agency works on whether a language model reaches for your business when somebody asks it a question in your category, and whether it describes you correctly when it answers. The job is deciding what makes a company citable, producing that evidence, and then checking whether the systems actually picked it up.

Four bodies of work sit inside that, and they run together instead of in sequence.

1. Diagnosis. Which questions your buyers plausibly ask, what the assistants say back today, who gets named instead of you, and which sources those answers are built from.
2. Fixing what you control. How your company is described, how your pages are constructed, and whether a crawler can read them at all.
3. Producing what you do not have yet. Original data, named expert positions, and accurate coverage on sources outside your own domain.
4. Re-measuring. The same question set run again, so movement is visible instead of asserted.

Beyond execution, an agency owes you honesty about the ceiling of this work. Nobody controls what a model says. There is no submission form, no bid, no support queue that gets a brand inserted into an answer. What can be influenced is the evidence a system has available when it composes one, and that is where every legitimate lever sits. A firm offering you guaranteed citations is describing a mechanism that does not exist.

Brand Vision has been building brands, sites and search programs since 2018, across more than 500 projects, and our rating is 5.0 on Clutch from 64 or more verified client interviews. That history matters here for an unglamorous reason. Most of what earns a citation is the same durable work that earns a ranking, held to a stricter standard of clarity. The named practitioners here are the people who do it, which counts for more in this discipline than in most, because attributed human expertise is itself a citable asset.

Where it belongs commercially. This is one track inside the wider search practice instead of a separate product with its own invoice, and the diagnosis usually arrives as part of a full site audit so you can read the AI picture against the organic one. Buying this alone, on a site with a broken foundation, is buying the last ten percent first.

Seven, sequenced instead of bundled, because three of them return nothing until the first four are done. Which ones you actually need comes out of the baseline, and on plenty of accounts the honest scope is three of the seven.

AI visibility auditing and prompt-level tracking. We build a question set that mirrors how your buyers really ask, run it across the major assistants, and record what comes back. Who gets named, what gets cited, how you are described when you appear, and where you are absent entirely. That baseline is the only thing that makes later movement legible.

Entity and knowledge-graph clarity. A knowledge graph is a machine-readable map of things and how they relate. This work makes sure every system holding facts about your company holds the same ones, so what you do, who you serve, where you operate and who runs it stop contradicting each other.

Content structuring for extraction and citation. Rebuilding priority pages so a single claim can be lifted out cleanly, with the question stated in plain language and answered in the sentences directly beneath it. Related to content strategy and a distinct task from it.

Original research and evidence production. Surveys, analysis of data only you hold, benchmarks, and named expert positions. The material that cannot be paraphrased out of existence.

Third-party source and citation building. Getting your business accurately described on the sources these systems lean on, meaning trade publications, directories, review platforms, community threads and reference sites. Shares machinery with link building and digital PR and optimizes for the accuracy of a mention as much as for the link.

Technical accessibility for AI crawlers. Whether the bots can fetch, parse and understand your pages. It overlaps with technical SEO and carries failure modes that standard audits miss, which is covered further down this page.

Measurement and reporting on share of AI mentions. Monthly, against the original question set, reported as share of appearance and citation instead of as a position.

What we will not sell you. Volume content generated by a model, which is the fastest way to look like every other source and get cited by none of them. Fabricated research. Review incentives that breach platform terms. And any claim that a specific prompt will name you by a specific date.

GEO stands for generative engine optimization and AEO stands for answer engine optimization. They describe one ambition under two labels, which is being present and accurately represented when a machine writes the answer instead of handing over a list of links. The labels are marketing. The work underneath them is real.

The overlap with search work comes first, because it is most of the story. A site that loads properly, states plainly what it does, and gets described the same way by people who do not work there tends to perform in both places. That is no accident. These systems were trained on the open web and many of them retrieve from a live index at question time, so the foundations of the organic search program carry across almost entirely. If a vendor pitches this as a discipline unrelated to your search foundations, you are being asked to fund the same work twice.

Where it genuinely diverges is worth knowing before you budget for it.

• The unit of competition is a passage instead of a page. A model lifts two or three sentences that answer the question. The quality of the rest of the page affects whether it trusts you as a source, and those sentences are what travels.
• Being described correctly matters as much as being visible. A system can name your company and get your service wrong, which a ranking report would score as a win.
• Consensus beats assertion. Nine independent sources agreeing on what you do counts for more than your own homepage saying it once, because anything resolving conflicting claims defaults to agreement.
• Volume stops paying. Forty thin pages can win a long tail in search. Only one description of your company gets used in a generated answer, so the marginal page earns nothing.
• The scoreboard changes shape. You track how often you appear across a defined set of questions instead of where you sit in a list.

One difference matters especially in business-to-business buying, where the research phase is long and mostly invisible. An assistant now sits inside that phase, assembling a shortlist before anyone fills in a form. None of that appears in a traffic report, which is part of why this work is hard to sell and worth doing anyway.

Some of it is going, and the losses concentrate in one type of query instead of spreading evenly across your site. Questions with a factual answer now get answered in place, and the click that used to follow often does not happen. The useful exercise is working out how much of your traffic is that kind of query, because for most businesses it is a smaller share of the commercially valuable half than the headlines suggest.

The compression is concentrated on definitions, quick comparisons, how-to steps, hours, specifications and unit conversions, meaning anything somebody wanted as a fact and not as a destination. If a large part of your traffic arrives on informational articles built to catch a query and then hand a reader to a newsletter, that model is under real pressure and no amount of optimization reverses it.

Clicks hold up where the person still has to arrive somewhere. Transactional intent survives because a buyer has to land on a checkout. Local intent survives because somebody needs a verified phone number, a route and opening hours, which is why local SEO has kept its value better than most of the discipline. Considered purchases survive because a serious buyer reads three vendors properly before contacting any of them, and so does anything needing an account, a quote, a booking or a human. A product catalogue still gets visited because the transaction lives there.

Then the part that surprises people. What survives tends to convert at a higher rate, because the visitors who still click are the ones the answer did not satisfy, and an unsatisfied reader is usually further along. Total sessions falling while leads hold flat is a pattern we now see often enough to warn clients about in advance, and it is a reason reporting has to sit on leads, pipeline and revenue instead of on traffic counts.

The honest complication is attribution. Traffic arriving from an assistant frequently lands in your analytics as direct or carries no referrer at all, so the measured referral volume from these platforms understates reality on most accounts. Nobody has solved that cleanly yet. Where volume is needed inside the next quarter while all of this settles, paid search is the honest place to buy it, and funding both at once is a common answer instead of a compromise.

It is priced as a track inside a monthly search program instead of as a standalone project, because the durable parts compound and a one-time push decays. A one-off baseline and remediation can be bought on its own as a fixed scope, and for some companies that is the correct purchase.

How the work sequences, with the durations we see in practice.

1. Baseline and diagnosis, two to three weeks. The question set gets built, run, recorded and read against your competitors, and the technical and entity problems get inventoried.
2. Entity and technical remediation, weeks three to eight. Description consistency, structured data, rendering, crawler access. The fastest returns in the whole discipline sit here, and they are unglamorous.
3. Structural content work, month two onward. Priority pages rewritten so claims can be extracted, then extended across templates.
4. Evidence production, month three onward and continuous. Research, expert positions, and third-party coverage. The slowest track and the one that still matters in two years.
5. Re-measurement, first re-run at eight to twelve weeks, then monthly. Early runs establish variance so later ones can be trusted.

The largest single thing that moves the number is whether you already own data worth publishing or we are producing it from nothing, because research carries real production cost and a rewrite does not. Site size and condition come next, since correcting how a fifty-page brochure business is described costs a fraction of what the same work takes across a large estate holding a decade of inconsistent records. Then how contradictory your current footprint is, because a business that has changed names twice needs real hours of cleanup. Then how many assistants, languages and locations get tracked. And finally how much of the execution sits with your team instead of ours, which is the lever most clients underuse.

On timing, the same rule applies here as everywhere else in search. Meaningful movement takes four months to a year, and anybody promising first-page or first-answer results inside thirty days is either misleading you or working on questions nobody asks. Entity and technical fixes can show up faster, sometimes inside weeks, because you are correcting a machine's understanding instead of earning a position.

We work to a six-month minimum on ongoing programs, and you own every account, property and file involved. Roadmaps and diagnostics are priced as fixed-scope work and usually start with the diagnostic that comes first. Put your category in front of us and we will tell you plainly whether this is worth funding yet.

For a lot of companies the honest answer is not yet, and for a few it is already late. What separates them is whether the foundation underneath is sound, because this work amplifies a clear business and cannot manufacture one.

Buy it now if. You sell something people research before purchasing. Your category already returns assistant answers that name specific vendors, which takes ten minutes to check. Your site is technically healthy. And you have either data worth publishing or people willing to be quoted by name. Those four together are the profile where this pays.

Buy the smaller version if. You want to know where you stand before committing to a program. A one-time baseline plus entity and crawler cleanup is a legitimate scope, it takes a few weeks, and it frequently finds problems worth fixing regardless of what you decide next. Plenty of clients stop there for two quarters and that is a reasonable decision, not a failure of nerve. A scoped consultation is the cheaper way in.

Do not buy it yet if. Your pages cannot be rendered or crawled, in which case the fix is technical and comes first. You have no organic presence at all, since a business no source describes has nothing for a model to assemble. Your positioning is unsettled, because we cannot make sources agree on a description you have not chosen. Or your buyers do not research online, in which case the broader marketing plan has better uses for the money and we will say so.

The purchases that look like this and are not. Wanting a large volume of content written by a model, which is a different thing entirely and works against you here. Wanting a chatbot on your website. Wanting your own internal documents searchable by an assistant. All three are real projects and none of them is this one, and it is worth naming because they get quoted interchangeably.

The uncomfortable summary is that this discipline is about three years old as a paid service, the platforms rewrite the rules without notice, and the interventions with the strongest evidence behind them are the same ones a good search program was already doing. If you would rather spend the budget on an editorial plan and better technical foundations, you will capture most of the same benefit. We would rather tell you that than sell a fashionable line item.

A model does not look your company up in a file, it assembles you out of everything it has read. Every source describing your business is effectively a vote, and where sources contradict each other the version with the most agreement wins.

An entity, in this context, means the thing itself instead of the words used for it. Your company is an entity, so is each location, each product, each named person. What has to agree across every source that mentions you.

• The legal name, the trading name, and the variants people actually type, including the misspellings
• What you sell, described in the words your market uses instead of your internal product names
• Every location with its address, phone number and hours matching character for character across profiles, which is the same discipline that drives local search visibility
• Founding date, size, ownership, and any former names left behind by a rebrand or acquisition
• The people, with consistent titles and biographies, since expertise attaches to named humans and not to companies
• Category and classification fields on the platforms that use them, chosen deliberately instead of accepted as a default
• Structured data on your own site stating all of the above in machine-readable form, with the property that points at your external profiles so a system can confirm the connection

The common failure is boring. A company rebrands, moves office, or absorbs another business, and four years of records now describe two slightly different organizations. Old addresses persist on directories nobody remembers creating. A founder's title differs on three platforms. None of it looks urgent and together it makes you an unreliable entity, which is a bad thing to be when a system is deciding whether to state something about you with confidence.

Fixing it is cleanup work with a claim register behind it, meaning one document that defines every fact about the business and what the approved wording is, then propagation of that wording everywhere. Our own Toronto office details are handled the same way for the same reason.

Where the description itself is still unsettled, this stops being a data problem. If the leadership team gives three different answers about what the company does, no amount of markup fixes it, and the honest first step is positioning research before anything gets published.

Your homepage is the least persuasive source of information about your company in existence, because every homepage claims to be the leading one. A system built to weigh evidence discounts self-description and looks for agreement among parties with nothing to gain from your success.

Ordered from weakest to strongest as evidence, which is roughly the reverse of where most marketing budgets go.

1. Your own marketing pages. Necessary, and read as a claim rather than a fact. They establish what you say about yourself and settle almost nothing.
2. Your own research and documentation. Stronger, because a number with a stated method can be checked by whoever reads it.
3. Profiles and directories you fill in yourself. Weak as persuasion, genuinely useful for confirming that the entity exists and is consistent.
4. Customer reviews at volume. Independent, specific, and frequently quoted back in generated answers almost verbatim.
5. Unpaid community discussion. Forums, subreddits, industry Slack groups, question sites. Heavily represented in training data and impossible to fake convincingly at any scale.
6. Trade and independent editorial. Somebody with a professional reputation at stake describing what you do. This is where earned coverage and AI visibility become the same investment.
7. Independent testing, awards with real criteria, and citation by other recognized experts. The strongest and the slowest, and largely outside anybody's direct control.

The uncomfortable part is that the strongest sources are the ones we cannot buy. What we can do is identify which sources your category's answers actually cite, find the genuine reasons for those sources to mention you, and make sure that when they do, the description is accurate. Correcting a wrong fact in a source that already cites you is often worth more than a new mention.

What we refuse. Fabricated reviews, incentives that breach a platform's terms, and seeded community posts pretending to be customers. Those get detected, they carry a reputational cost far larger than the gain, and in regulated categories they create a compliance problem on top of it. This matters most for law firms and health and wellness brands, where the platforms and the regulators are both watching what gets said about outcomes.

A model can only quote what it can isolate, so the practical test is whether a single paragraph, read with nothing around it, still answers a question completely. Most business writing fails that test because it was built to be read top to bottom by a person who already arrived.

What we check, page by page, on the pages that matter.

• The heading is the question, phrased the way somebody would actually ask it, instead of a clever two-word label
• The answer arrives in the first sentence or two underneath, before context, history or qualification
• Each passage stands alone, with no back-references to an earlier section and no pronouns pointing at a subject named three paragraphs ago
• One claim per paragraph, so lifting it does not drag in a second unrelated idea
• Numbers, dates, units, currencies and caveats stated inline, in text, next to the claim they qualify
• The company or product named again instead of replaced with the word "it", since a lifted sentence loses whatever antecedent it depended on
• Plain nouns on first use, with internal product names glossed, because a model matching a question to a passage matches vocabulary
• Definitions before elaboration, so the paragraph a system reaches for is the one that defines the thing
• Real lists formatted as lists, since a promise of five steps followed by prose reads as one undifferentiated block
• Nothing load-bearing living only inside an image, a video, or a PDF that parses badly

Structured data helps a machine understand what a passage is, and it does not make a weak passage worth quoting. The writing carries the weight. This is also why the site-wide FAQ format has held up so well through every platform change. A question with a direct answer under it was extractable before anybody used the word GEO.

There are honest costs to writing this way. A page built for extraction reads more like a reference document and less like a persuasive one, which is a trade worth making on documentation and definitional pages and worth refusing on your homepage and campaign pages. The second cost is scale, because across large site estates this cannot be hand-applied. It becomes a template and an editorial standard, or it reaches forty pages and stops.

Original data and named human expertise are the only two things a model cannot restate without pointing back at you. Everything else gets compressed into a sentence that credits nobody, because the same information exists in four hundred other places.

That is the whole mechanism. A system summarizing consensus has no reason to name a source. A system reporting a specific number, held by one organization, has no way to use it without saying whose it is. Attribution follows scarcity.

Proprietary data. Something only your operations produce, aggregated and anonymized into a figure your market wants. Transaction patterns, pricing movement, failure rates, cycle times. Most companies are sitting on this and have never published any of it.

An original survey. Fielded properly, with sample size, population and method stated so it can be judged. Around 400 responses puts margin of error near five points, which is a field convention and not a Brand Vision figure. A survey with an undisclosed method gets ignored by the careful sources and those are the ones worth having.

Benchmarks and reference tables. Numbers other people need repeatedly and cannot compile themselves. These earn citations for years and quietly become the thing your category argues about.

Named expert positions. A person with a title, a track record and a real opinion, saying something specific enough to be wrong. Anonymous corporate hedging is unquotable by design. This is one reason who we hire matters commercially as well as culturally.

Documented outcomes. Real cases with the numbers, the timeframe and the constraints stated, including what did not work.

The honest problem is cost. Research is slow, it needs someone competent to design it, and it can conclude something inconvenient. Publishing twelve articles instead is faster and cheaper, which is exactly why the articles earn nothing. Companies in technology and SaaS have the easiest path here, because usage data already exists and only needs permission and a careful hand. For everyone else it starts with one honest question about what your business knows that nobody outside it can find out.

On a meaningful number of modern builds they cannot, and the reason is that several major AI crawlers fetch a page's JavaScript and never execute it. Anything the browser assembles after load simply does not exist for them, so a beautifully built site can be functionally blank to the systems you are trying to reach.

The checks that find it, in the order we run them.

1. Rendering. Look at the raw HTML a server returns with scripts switched off. If the body is an empty container waiting for a framework, your content is invisible to any crawler that does not render. Server-side rendering, static generation, or prerendering for bots fixes it, and the fix belongs with whoever owns your codebase. Platforms that ship real HTML by default, Webflow among them, never had this problem.
2. Bot access at the edge. Firewalls, bot management and rate limiting block AI agents by default on a lot of stacks, and nobody inside the company can see it happening. Log files settle the argument in an afternoon.
3. Crawler directives in robots.txt. The agents that identify themselves can be allowed or blocked by name, and the list changes often enough to need reviewing instead of setting once. The decision is less obvious than it looks, because blocking a crawler that gathers training data can also remove you from the live retrieval that answers questions on some platforms. We lay out both sides and you decide.
4. llms.txt, which is an open question and not a settled standard. The proposal is a plain text file at your root telling a model which pages matter and how to read your site. It costs almost nothing to publish and it may well become useful. What it is not, today, is something the major platforms have committed to reading, so we add one and make no claim about its effect until a server log shows something fetching it.
5. Content behind an interaction. Tabs, accordions, modals, infinite scroll and anything requiring a click. If a person has to act to reveal it, assume it was not read.
6. Speed and timeouts. Bots abandon slow responses sooner than people do, which is one of several places this work and crawl and rendering work are the same task under two names.

None of this is exotic. It is the reason we look at delivery before strategy on every engagement.

Measurement here is share of appearance across a defined set of questions, sampled repeatedly, and it behaves like a poll instead of a scoreboard. A generated answer has no fourth position for you to occupy, so the ranking metaphor collapses the moment you try to use it.

How it works in practice. We agree a question set with you, usually somewhere between fifty and a few hundred prompts, covering category questions, direct comparisons against named competitors, problem-shaped questions your buyers ask before they know the category, and location variants where those matter. That set is fixed, so later runs are comparable. Then it gets run on a schedule across the major assistants and the results get recorded.

What comes back in the monthly report.

• Share of appearance, meaning the percentage of the question set where you are named at all
• Citation share, meaning which domains get linked or quoted in those answers, which tells you where the next piece of work belongs
• Description accuracy, because being named and mischaracterized is a distinct problem that deserves its own line
• The competitor picture across the identical set, including who is gaining
• Referral traffic from assistant platforms, carried with the caveat that attribution from these sources is unreliable and understates itself

The limits, stated plainly because they are real. These systems are non-deterministic, so the same prompt asked twice can return different answers. Results shift with personalization, location, account history and model version, and a platform can change behaviour overnight with no announcement. A single observation therefore proves nothing, which is why we sample the whole set repeatedly and report a rate with its variance instead of a screenshot. Anybody showing you one flattering answer as evidence of a program working is showing you noise.

Two commitments hold this together. Reporting connects to leads, qualified pipeline and revenue, never to impressions or mention counts presented as an outcome. And nobody here will guarantee a citation, because no method available can carry that promise. What we will tell you is whether your share of appearance is moving, what moved it, and what did nothing.

If you want the current picture before committing to anything, a one-off read of your category will show you what the assistants say today. Ask us what it looks like and we will run a sample set before you spend a dollar on changing it.