Audit before you integrate: what AI already says about your brand
Most conversations about AI in marketing start with adoption. Which model, which workflow, which agent. The question that should come first is quieter and more uncomfortable: what do these systems already say about your organisation when a buyer asks?
Every brand now has a second description of itself, written by models that have read a mixture of its own pages, press coverage, forum threads, analyst notes and pages it has never seen. That description is being delivered to buyers in answers rather than links. Integrating AI into marketing before auditing that description is building the roof before the walls.
The audit most teams skip
In an October 2026 guide, the AI-visibility firm Wellows frames brand presence in model answers as three distinct checks. Presence asks whether the brand is named in a relevant answer at all. Proof asks whether the brand's own pages are cited as sources. Portrayal asks whether the description is accurate. A brand can pass one and fail the others. Being named often while being described wrongly is a worse position than being absent, because the wrong description travels under the brand's name.
Wellows argues that a wrong answer about your own brand does more damage than a missing one, because what it calls a shadow reputation inside these answers is pieced together from pages the brand did not write. That is the practical case for auditing before integrating. You cannot brief a model, a workflow or an agency against a baseline you have never measured.
Why a single check is not enough
The same guide recommends running 25 to 40 real buyer prompts across ChatGPT, Gemini, Perplexity, Google AI Overviews and AI Mode, in a logged-out window, recording per answer whether the brand is named, whether a URL is cited, the position in the response and the tone. It also recommends repeating the exercise weekly for a month, because any single round is unreliable.
That cadence is vendor-stated and worth treating as a working assumption rather than a law. The underlying reason it matters is mechanical. Model answers are assembled from retrieval plus generation, and both sides shift. A prompt asked twice in the same week can return different cited domains, different ordering and a different summary sentence. A one-off screenshot of ChatGPT saying something flattering, or something wrong, tells a strategist almost nothing about the steady state. A month of weekly observations begins to show which claims are stable, which flicker and which depend on a single thin source the brand could influence.
Sources are local
Wellows reports its own analysis of 20.5 million scored AI citations collected between 8 March and 31 July 2026. Across that dataset, 301 domains with Domain Authority 90 or above received 17 per cent of citations, while 386,000 distinct domains appeared as cited sources across 42 markets. The figures are the firm's own and have not been independently replicated, so they are a signal rather than a settled benchmark.
The operational point stands either way. Citation sources are heavily local. A UK strategist cannot read a US citation list and assume the same titles show up in a London buyer's answers. If the audit is going to inform where the brand earns coverage, where its executives write and which third-party pages it tries to correct, it has to be run market by market. A single global snapshot will miss the specific publications that actually feed answers for UK buyers.
The accuracy problem is real but narrower than the headline
Wellows also cites a Columbia Journalism Review Tow Center test of eight AI search engines using 1,600 queries, which found that more than 60 per cent of searches received an inaccurate response, and that over half of Gemini's responses contained made-up or broken hyperlinks. Wellows itself flags that this test used news queries rather than brand or category queries, so it is a warning about the medium rather than a measurement of how often models get a specific company wrong.
Treated honestly, that is still enough to justify the audit. If a system that answers news questions is wrong a majority of the time, a strategist has no basis for assuming it is reliably right about a mid-sized B2B brand with a thinner public footprint. The null hypothesis should be that the description is partly wrong until a logged audit says otherwise.
What this changes for the integration plan
For a senior strategist weighing AI adoption, the sequence matters more than the tool choice. An audit changes what the integration is trying to do. If the brand is named often but described inaccurately, the first job is correction: rewriting the pages that models retrieve, tightening the executive bios and leader essays that supply the phrasing, and placing compatible language in the external pieces the models are already citing. If the brand is largely absent, the first job is presence: earning coverage and publishing substantive pages in the categories where buyers actually ask. If citations are dominated by domains the brand cannot easily influence, the job becomes selecting the two or three it can.
None of that requires an AI workflow to start. It requires a baseline. The baseline is cheap to build, uncomfortable to read, and the only honest input for an integration plan.
A quiet next move
The practical move for a strategist this quarter is narrow. Pick 30 prompts a real buyer would type. Run them across the five surfaces Wellows lists, in a logged-out window, once a week for four weeks, in each market that matters. Log presence, citation, position and tone per answer. Read the result before signing off on any AI integration budget. The integration plan written after that audit will not look like the one written before it, and the gap between the two is the point.
Source: Wellows, Audit brand visibility on LLMs, 12 October 2026. https://wellows.com/blog/audit-brand-visibility-on-llms/