aijelen.hu / Articles / Guide

How to Record an AI Visibility Baseline Before Changes

Establish a frozen snapshot of your brand's current AI presence across platforms to enable accurate measurement of future visibility shifts.

By Zoltán Varga2026-11-054 min read

Why Baseline Comes First

You cannot measure improvement without a fixed starting point. AI systems generate responses based on training data and live retrieval, creating two distinct layers of information. A baseline captures the state of both layers before you alter any content or structure. Without this record, you cannot distinguish between a model update and the effect of your specific interventions.

The Per-Entity Bias Mapping framework identifies three failure modes that require different solutions. Type A entities are invisible across all queries. Type B entities appear but generate incorrect facts due to low verification rates. Type C entities suffer from infrastructure gaps where knowledge graphs lack anchors. Your baseline audit determines which mode applies to your brand.

Recording this state prevents the confusion of attributing random model drift to your work. If you launch a new website without a baseline, you cannot prove the new site caused better answers. The baseline provides the evidence needed to link specific actions, like adding structured data, to changes in AI outputs. It turns subjective impressions into measurable data points.

What to Capture in the Audit

You must capture five specific query types that cover identity, function, location, products, and personnel. These queries include the company name, the question of what the company does, regional industry comparisons, product combinations, and executive roles. Every response must be recorded for accuracy and completeness. You note whether the AI mentions the brand, if the facts are correct, and if hallucinations occur.

The audit distinguishes between parametric and retrieval platforms to measure knowledge lag. Parametric models rely on static training data, while retrieval systems access live web sources. You compare the answers from a closed-book model against a retrieval-augmented one. A discrepancy indicates a temporal gap where the retrieval layer knows a fact the parametric layer does not.

You also inventory the technical infrastructure supporting the entity. This includes checking for a Wikidata QID, a Wikipedia page, and correct schema.org markup on the website. You verify that social media profiles link back to the official domain. The absence of these elements explains why an entity fails to appear in search results despite having relevant content online.

Platforms and Dates Matter

You select four primary platforms for the baseline snapshot to ensure broad coverage. ChatGPT serves as the pure parametric control when browsing is disabled. Perplexity acts as the primary retrieval-augmented system. Gemini and Copilot provide blended models that mix both approaches. You run the same queries on all four platforms on the same day to eliminate time-based variables.

The date of the snapshot is critical for calculating lag metrics. If a leadership change occurred six months ago, the baseline reveals how many platforms still report the former leader. You calculate the difference between the event date and the current model knowledge. This delta score quantifies the severity of the lag and guides your remediation strategy.

You save the output in two formats: a raw text log and a screenshot. Text logs allow for automated comparison later, while screenshots preserve the visual context of the interface. You record the exact timestamp of each query. This documentation ensures that future audits can replicate the conditions precisely for valid comparison.

Freezing the Query Set

You define a fixed list of queries before running any tests and never change them during the project. Changing the questions introduces noise that invalidates the comparison. The set must remain constant for the duration of the baseline and subsequent tracking periods. This consistency allows you to detect subtle shifts in answer quality over time.

Each query must target a specific fact rather than a general opinion. Questions like "Who is the CEO?" yield clear true or false answers. Vague questions like "How good is the service?" produce variable responses that are hard to measure. You avoid leading questions that suggest the answer within the prompt itself.

You document the exact wording used for every query. Even minor phrasing differences can trigger different retrieval paths or model behaviors. You store the query list in a central document accessible to the entire team. This protocol ensures that every member of the project runs the same tests, maintaining data integrity across the workflow.

Storing the Baseline Data

You organize the collected data into a structured format that links queries to platforms and outcomes. Each row in your log represents a single query-platform pair. Columns capture the raw output, the accuracy status, and any noted hallucinations. You attach the corresponding screenshot file to each entry. This structure enables quick filtering and analysis later.

You calculate a lag score for each entity based on the comparison between parametric and retrieval results. A score of zero means both systems agree. Higher scores indicate significant divergence where the retrieval layer is current but the parametric layer is outdated. This score becomes the primary metric for tracking progress over the twelve-week sprint.

You treat the baseline as a living document that updates only with formal audits. You do not modify the initial snapshot to account for new information. The original record remains the reference point for all future measurements. This discipline ensures that your claims about visibility improvements are backed by immutable evidence rather than memory or assumption.

AI-assisted: drafted with a language model from the author's research notes, then edited and checked before publication. Author: Zoltán Varga, AI visibility researcher and AI ecosystem architect, Budapest.