For the past couple of months, I have been in the trenches doing AI visibility work with live clients.
That has meant evaluating measurement platforms, designing prompt sets, building reporting structures, reviewing daily results, analyzing competitors, tracing citations, and trying to understand what is actually driving the answers produced by ChatGPT, Claude, Gemini, Perplexity, and other AI platforms. The experience has made one thing increasingly clear: AI visibility is becoming a real marketing discipline.
The opportunity is significant, but the work is frequently oversimplified. AI visibility is often presented as a software workflow where a company enters its URL, runs some prompts, receives a score, and follows a list of recommendations.
The platforms are useful, but they are only the measurement layer. The dashboard is not the strategy. It is where the real work begins.
The first challenge is deciding what to measure
Before a company can improve its AI visibility, it has to define what it wants to be visible for.
Which audiences matter? Which problems should lead an AI platform to recommend the brand? Which use cases are commercially important, and which competitors should be included in the analysis?
Those questions sound straightforward, but answering them requires real strategic choices. A company might want to be visible for its category, specific customer problems, product capabilities, comparison questions, implementation guidance, or buying recommendations.
Trying to measure everything at once usually produces a noisy prompt set and an unhelpful score. The better approach is to identify the moments in which the brand should reasonably appear and organize prompts around those moments.
Prompt design also affects the results more than many people realize. A prompt that names the company is fundamentally different from an unbranded prompt asking for the best solution to a buyer's problem.
Branded and unbranded prompts should not be treated as equivalent measures of visibility. Neither should informational questions, product comparisons, provider recommendations, implementation questions, and category education prompts.
The reporting structure has to preserve those distinctions. Otherwise, a high branded visibility score can obscure the fact that a company is largely absent when prospective customers are asking unbranded discovery and recommendation questions.
Measurement design is not administrative setup. It determines whether the baseline will produce useful business insight or simply generate another dashboard.
One day of results is not a baseline
A single visibility score after one day of runs does not tell a company how it is performing.
The same prompt can produce different responses on different days. The brands mentioned, sources cited, claims emphasized, and recommendations made can all change.
That variation can come from changes in retrieval, differences in available sources, normal variability in generative responses, and ongoing updates to the models and systems around them. It does not make AI visibility impossible to measure, but it does mean the measurement system has to account for volatility.
A useful baseline requires repeated runs over time. It also requires looking below the aggregate score to understand whether a pattern is persistent or the product of a single unusual result.
The goal is not to react to every daily movement. The goal is to identify sustained strengths, recurring gaps, meaningful competitive changes, and factual problems that continue to appear.
This is why AI visibility has to be tracked and managed over time. A one-time audit can provide a snapshot, but it cannot tell a company whether its position is stable, improving, declining, or simply fluctuating within a normal range.
Different models tell different stories
Another clear learning is that there is no single AI visibility result.
A company can perform well in ChatGPT and poorly in Claude. Perplexity may rely heavily on one set of sources, while Gemini surfaces a different group of brands and citations.
The models may also interpret the same company differently. One may understand the product category correctly but miss an important capability, while another may recommend the company for a use case it does not actually support.
That makes model coverage a significant measurement decision. Many entry-level AI visibility packages we evaluated did not include Claude, even though Claude frequently produced materially different results from the other major platforms.
Excluding a model can make a program less expensive and easier to operate. It can also create an incomplete view of how buyers using AI platforms may encounter the brand.
The answer is not necessarily to average all model results into one number. Averaging can hide the most valuable information, including where a company is strong, where it is absent, and where the models disagree about basic facts.
AI visibility reporting should preserve the model-level view. Each model is effectively a distinct channel with its own sources, behaviors, strengths, and gaps.
Best practices are not the same as a diagnosis
There are useful AI visibility best practices.
Technical accessibility matters. Structured content matters. Consistent product information, clear page architecture, strong third-party evidence, and content that answers specific questions all remain important.
But generic best practices only get a company to the starting line. They do not explain why a specific competitor appears in a recommendation, why one model misunderstands the product, or why an important customer use case never triggers the brand.
The market is also changing too quickly for a six-month-old GEO checklist to function as a complete strategy. Some recommendations remain useful, while others may be outdated, irrelevant to the company, or counterproductive when applied without understanding the underlying problem.
Publishing more content is a good example. More content can create additional opportunities for retrieval, but it can also amplify confusion when product descriptions, claims, positioning, and terminology are inconsistent across a company's website, support documentation, partner pages, case studies, and external coverage.
The real impact comes from digging into the actual prompt results.
Why did a competitor appear? Which source influenced the answer? What language did the model use to describe the category? Did the citation actually support the claim being made?
Those questions lead to specific actions. A company may need a clearer product page, stronger comparison content, updated partner documentation, better evidence for a claim, a technical fix, or a more consistent explanation of how the product works.
A generic checklist cannot identify those priorities. Prompt-level analysis can.
Visibility and accuracy are separate outcomes
One of the most important lessons from the work so far is that a company can have strong AI visibility while still being represented incorrectly.
We recently saw a brand performing well in the results, but the model was explaining how its product worked incorrectly. From a visibility-score perspective, the brand was succeeding.
From a business perspective, the result was a problem.
A prospective customer could read the answer and walk away with the wrong understanding of the product, its capabilities, or how it could be used. The company was visible, but the visibility was reinforcing misinformation.
This means factual accuracy has to be measured separately from mentions, recommendations, rankings, and citations. A model can cite a source without the source supporting the specific claim made in the response.
It can also combine information from multiple sources and reach a conclusion that no individual source actually states. Citation presence is useful, but it is not a substitute for claim-level validation.
Companies need a process for reviewing how AI platforms describe their products, capabilities, customers, limitations, activation paths, pricing, and competitive position. Visibility without accuracy is not a successful outcome.
AI visibility is becoming an operating discipline
The more time I spend doing this work, the less it looks like a one-time SEO project.
AI visibility requires measurement design, prompt governance, competitive analysis, content strategy, technical execution, citation development, product marketing, and factual quality assurance. It also requires coordination across teams that may not be accustomed to working from the same source of truth.
The work follows an ongoing cycle.
First, define the questions and audiences that matter. Then establish a multi-model baseline, inspect the underlying responses, identify the causes of performance, prioritize changes, implement them, and measure the results again.
That process has to distinguish persistent patterns from daily noise. It also has to account for changes in the models, sources, competitors, and ways customers use AI platforms.
The software helps organize and surface the data. Human judgment is still required to determine what should be measured, what the results mean, which changes are worth making, and whether those changes actually improved the customer experience.
The goal is not simply to increase a score
After a couple of months in the trenches, my view of AI visibility is more optimistic, not less.
Brands can influence how they are discovered, described, cited, and recommended by AI platforms. But achieving that requires more than subscribing to a tool or applying a standard set of GEO tactics.
It requires treating AI visibility as an ongoing marketing discipline. The organizations that do this well will build measurement systems they trust, understand why competitors are winning, improve the information available to the models, and validate that their brands are being represented accurately.
At OuterCurrent, that is how we are approaching the work. The objective is not simply to make a visibility score go up.
The objective is to help brands appear in the right conversations, for the right reasons, with information that is clear, useful, and factually correct.