How Optimly Works

The simplest version of an AI brand audit is: type your company name into ChatGPT, read the response, decide if it's right.

That approach has three fundamental problems.

First, AI outputs are stochastic. Ask ChatGPT the same question twice and you may get different answers. Ask it with slightly different phrasing and you'll almost certainly get different answers. A single prompt-response pair tells you what the model said once, in one context, for one framing. It doesn't tell you how the model represents your brand across the range of real buyer interactions — which vary in phrasing, intent, competitive context, and specificity.

Second, different models have different representations. ChatGPT, Claude, Gemini, Perplexity, and Grok each have different training data, different retrieval architectures, different source weighting, and different parametric knowledge. A brand can be accurately represented in Claude and completely misrepresented in Gemini. A single-model audit misses the cross-model variance that determines how buyers across different AI tools experience your brand.

Third, knowing what's wrong doesn't tell you why — or how to fix it. If ChatGPT miscategorizes your company, the useful question isn't "what did it say?" It's "what caused it to believe that, and what would change its belief?" Without causal attribution, remediation is guesswork — you publish content and hope the model notices. That's not engineering. It's prayer.

Optimly's technical approach addresses all three problems.

Before measuring accuracy, you need a reference point. Optimly's Define phase establishes a machine-readable brand ground truth: the canonical representation of what AI models should believe and communicate about a company.

This isn't a messaging document. It's a structured knowledge specification that includes:

Category definition. The precise category the brand belongs in, including parent categories, adjacent categories it should not be confused with, and the semantic boundaries between them. This matters because AI models infer category membership from distributed signals — and the boundaries between similar categories are often ambiguous in training data.

Entity disambiguation. Explicit identification of entities the brand should be distinguished from — competitors with similar names, products in adjacent categories, the brand's own historical positioning that may differ from current positioning. This becomes the reference set for measuring confusion and conflation in model outputs.

Capability mapping. Structured representation of the brand's actual capabilities, use cases, and target segments — with version awareness. Models often carry knowledge of deprecated capabilities. The ground truth specifies what's current and what's historical so accuracy measurement can distinguish "the model knows about us" from "the model knows an outdated version of us."

Competitive positioning. Where the brand sits relative to its actual competitive set — not the competitive set the brand wants to be in, but the one that accurately reflects market reality. This prevents the brand-reality gap that undermines positioning in AI systems.

The ground truth becomes the scoring rubric. Every subsequent measurement is "how close is the model's representation to this specification?"

Detection is where Optimly's technical approach diverges most from standard AI monitoring.

Prompt distribution design. Rather than running a fixed set of prompts, Optimly designs prompt distributions that mirror real buyer behavior. This includes varying query framing (direct brand queries, category queries, comparative queries, use-case queries), buyer intent (research, evaluation, shortlisting, objection handling), competitive context (head-to-head comparisons, category recommendations, alternative suggestions), and specificity (broad category questions vs. targeted capability questions).

The goal is to characterize the model's brand representation across the full space of buyer interactions — not just the handful of prompts a human would think to test.

Stochastic output evaluation. For each prompt class, Optimly runs multiple evaluations to capture output variance. This produces a distribution of responses rather than a single answer, enabling statistical analysis of representation consistency. A brand that's accurately described 90% of the time but miscategorized 10% of the time has a different problem than a brand that's consistently miscategorized — and the fix is different.

Cross-model comparative analysis. Every evaluation runs across all five major models independently. The cross-model view reveals which models are aligned with ground truth, which diverge, and how they diverge differently. This is critical because each model's misrepresentation typically has different causal sources — a fix that corrects ChatGPT's representation may not affect Claude's.

Layer separation. Where model architecture allows, Optimly separates parametric responses (model answering from trained knowledge only) from retrieval-augmented responses (model answering with web search or tool use). The gap between these two response types reveals whether the problem is in the model's foundational beliefs, in the content available for retrieval, or in the interaction between the two — each requiring different remediation strategies.

Scoring. The Brand Authority Index is computed from this distributional analysis — not from individual prompt tests. The 0-10 score reflects accuracy across prompt distributions, across models, and across the seven diagnostic dimensions (categorization, capabilities, competitive positioning, target market, differentiation, currency, and factual accuracy). This makes the BAI a statistically grounded metric rather than a subjective rating.

Knowing that a model misrepresents your brand is useful. Knowing why is what makes the problem fixable.

Source mapping. For each detected misrepresentation, Optimly traces the causal chain back to the source content that's teaching the model the wrong thing. This involves analyzing the corpus of web content that mentions the brand — directory listings, review sites, press coverage, competitor comparison pages, documentation, historical content, social mentions — and mapping which sources carry disproportionate weight in model training and retrieval.

Influence weighting. Not all sources are equal. A directory listing on a high-authority industry site influences model training more than a blog comment. A well-structured comparison page with schema markup gets retrieved more reliably than an unstructured press release. Optimly's influence framework weights sources based on domain authority, content structure, retrieval ranking patterns, citation network position, and historical correlation with model output changes.

Causal attribution. The goal is to answer: "If we fix this specific source, what's the predicted impact on model representation?" This turns remediation from a volume exercise (publish more content everywhere) into a targeted engineering problem (fix these three sources for maximum impact). In practice, a small number of high-influence sources often account for the majority of misrepresentation — which means targeted fixes produce outsized results.

The Deploy phase executes fixes and verifies they work. This is where the measurement loop closes.

Causal source remediation. Based on influence attribution, Optimly identifies the specific sources to fix — and the specific changes needed at each source. This might mean correcting a directory categorization, updating a review site description, requesting corrections to press coverage, counterbalancing a competitor's comparison page framing, or addressing historical content that reflects outdated positioning. Fixes are prioritized by predicted impact on model behavior.

Knowledge engineering. For gaps where no high-influence source exists to fix — where the model simply lacks sufficient signal about the brand — Optimly creates structured content designed for model consumption. This isn't content marketing for humans. It's information engineering for AI systems: structured data, entity-rich content, format and placement optimized for training pipeline ingestion and retrieval system indexing.

Dual-timeline execution. Retrieval-layer fixes (improving what models find at query time) take effect within days to weeks as models recrawl content. Parametric fixes (changing what models believe at the training-data level) take effect over months as models incorporate new training data. Optimly runs both workstreams in parallel and tracks which layer is responding.

Verification through re-measurement. After fixes deploy, Optimly re-runs the same distributional evaluation against the same prompt distributions and scoring rubric. Did the Brand Authority Index move? Which models responded? Which diagnostic dimensions improved? Which didn't? This closed-loop measurement is what turns remediation from guesswork into engineering — every fix is a testable hypothesis with measurable outcomes.

Iteration. Fixes that moved scores get scaled. Fixes that didn't get redirected. New misrepresentation patterns that emerge (from model updates, competitive changes, or market shifts) get detected in the next evaluation cycle. Because AI models are continuously updated, AI brand reputation is a continuous process — not a one-time project.

The technical approach produces practical outcomes:

For marketing leaders: You get a quantified score (Brand Authority Index) that tells you exactly where your brand stands across every major AI model, what's wrong, and whether your investment in fixing it is producing measurable results. You can report to your board: "We improved our AI brand accuracy from 3.2 to 7.8 across models, and AI-referred pipeline increased 40%."

For technical evaluators: The methodology is grounded in statistical evaluation of stochastic systems, not subjective prompt testing. Source influence attribution provides causal explanations for misrepresentation. Cross-model analysis captures the full surface area of AI brand representation. And the closed-loop measurement framework makes every fix verifiable.

For the company overall: AI brand reputation becomes a managed, measurable function — not an invisible risk. You know what AI believes about you, why it believes it, whether it's changing, and whether your actions are causing the change.