Understand why AI engines recommend you or pass you over,
then fix it with evidence.

Scope the decisions to influence

The question panel reproduces how your clients choose a provider, not a list of keywords rewritten as questions. Each question is tied to a service, a client type, a decision stage (discovery, comparison, choice, validation), a language and a market. The wording comes from the field: questions asked at first meetings, contact forms, email enquiries, queries found in Search Console.

The panel is split into three sets. The core set is frozen so that one wave can be compared with the next. The exploratory set tests new questions without breaking the series. The monitoring set checks how AI engines describe your brand. We favour constrained questions that sound like a real decision: “which Franco-Thai law firm can handle a company acquisition in Bangkok?” rather than “best lawyer Bangkok”.

Measure presence under a written protocol

Under standard services, each question is asked twice on every included interface, in neutral, logged-out sessions without history. The two readings document presence and variability; if they disagree, the result is marked as variable rather than applying a fictitious “two out of three” majority rule. Interfaces are tested within a documented collection window (72 hours where that protocol applies), and answers are recorded within the selected service scope.

The interface panel is set per market and named in the protocol before collection: by default ChatGPT, Google AI Overviews / AI Mode and Gemini in Europe; ChatGPT, Gemini, Google AI Overviews and Meta AI in Thailand. For businesses serving Chinese-speaking clients, a custom-quoted module adds Doubao, Qwen, DeepSeek and Yuanbao, collected in Simplified Chinese. In each priority market, core questions are collected in the local language and in English: strong visibility in English can hide a complete absence in Thai or French. Every major model update is logged; if it makes two waves incomparable, we set a new baseline rather than splice two series together.

Results read at two levels. The KVS score summarises your presence and position on the agreed panel out of 100 (Presence 30, Position 20, Top positions 30, Coverage 20). Beneath the score, four rates are reported separately: presence, explicit recommendation, preference in comparisons and description accuracy. For a services business, the two that matter most are recommendation and preference in comparisons: being named is not enough, you need to be advised. Each answer is also recorded with the competitors named and the sources cited. The KVS score does not measure each of the nine KAIS AI Visibility Chain™ dimensions independently.

Build the brand fact sheet

Description accuracy can only be measured against a validated reference. Before assessing what AI engines say, we build a dated fact sheet with you (Brand Truth Document): official name and variants, addresses and locations, leaders and experts, services, markets served, positioning statements you are happy to see reused, claims that need a qualifier, and reference profiles.

The fact sheet feeds a correction queue: which third-party profiles are in scope, who files the request, and what “fixed” means. A correction is closed when AI engines describe you accurately in later readings, not when the request is sent. Namesakes, multi-location businesses and former names are handled at this step.

Trace each gap to a hypothesis

A measured gap is not yet a diagnosis. We attach each significant gap to the dimension of the KAIS AI Visibility Chain that best explains it, and test that hypothesis first. Absent, present but never recommended, named without your site as the source, misdescribed: four different problems, four different fixes. The symptom → dimension grid below is the starting point.

Each lead gets a priority score: gap severity × business importance × likelihood of influence ÷ effort. The formula puts a directory listing that takes a week to fix and a multi-quarter PR programme on the same page without treating them as equal tasks. We keep three priority hypotheses at most, each with an action, an owner (KAIS, your team or a third party), the expected indicator and a re-test date.

Make your pages accessible and extractable

A page that ranks well on Google is not automatically readable by an answer engine. We check that your priority pages are crawlable, indexable, correctly canonicalised and linked to each other; that key text sits in the served HTML rather than in an image, a PDF or a script; and that answer-engine bots can reach them. This is confirmed with server logs, not just configuration.

The decision to allow or block each bot is documented separately: training crawlers (GPTBot, Google-Extended, CCBot) do not play the same role as the agents that fetch a page while answering (OAI-SearchBot, ChatGPT-User, PerplexityBot). On the writing side, each section answers in its first sentence and names services, audiences, markets and conditions explicitly. SEO remains an amplifier, decisive on Google AI Overviews where organic rank still carries weight; on ChatGPT or Perplexity, a business can be named without being on Google's first page.

Produce what helps buyers choose

AI answers that steer a choice draw on decision-support content, not general definitions. Before creating or rewriting a page, we score it on six criteria: usefulness after the AI answer (will the reader still need to visit to check, compare or book?), citation potential, association with your brand, business value, proprietary advantage (data, expertise or access only you have) and effort.

For a services business, this means making explicit what engines cannot guess: the ideal client, the process, deliverables, timelines, engagement terms, price ranges when you are willing to publish them, and the next step. Regulated sectors add jurisdiction, credentials, fees, effective dates and mandatory disclosures. We separate the evidence page, the one the engine cites, from the destination page, where your future client books a meeting: they do different jobs.

We advise against pages that split one question into variants, definitions rewritten without real expertise and comparisons with no stated criteria.

Strengthen corroboration and consistency

Answer engines cross-check what you say about yourself against what independent sources say about you. We classify the sources they cite on your questions by role: your own pages; earned sources (specialist media, professional directories, comparisons, reviews); community spaces; and paid placements, tracked separately and always disclosed. An action that only pays off because an engine favours a format today is ruled out: we keep the ones that retain audience, reputation or relationship value even if cited sources change.

Your public profiles (Google Business Profile, LinkedIn, industry directories, Wikidata where the entity is eligible) are aligned with the fact sheet. On Wikipedia: we assess notability and improve independent sources; no article creation or edit is guaranteed, as editorial decisions belong to the Wikipedia community.

Re-test, then keep, expand, revise or stop

After each action, we check whether the expected indicator moved, on the same questions and under comparable conditions. A comparison page is judged on recommendation and preference in comparisons; an entity fix on accuracy; access and structure work on your site being displayed as a source. An action is kept, expanded, revised or stopped based on what the measurement shows, not because it was delivered.

In monitoring programmes, the core set is collected every month; cited sources, markets and hypotheses are reviewed every quarter; a re-test is triggered after a redesign, an offer change or a major model update; the full panel is reviewed every six to twelve months, with the core set kept.

Business impact is read through the Outcome dimensions of the Chain (Referred, Contacted) and reported by confidence level, never added up into a single figure. Observed: visits and enquiries from AI assistants in your analytics. Declared: answers to an optional question on your form (“Did you discover us through an AI assistant?”). Indirect: branded searches and direct enquiries. Modelled: a range with its assumptions, offered as decision support and not as attribution. AI referral traffic is still small compared with traditional search: what matters first is being selected, before the click.