Automated GEO Audits with Claude Cowork – Stradiji”Brand visibility in AI search isn’t a one-off audit, it’s a trend you have to watch week to week. The “paste these 15 prompts into Google AI Mode” audits going around are a decent start, but they have three gaps: they look at a single engine, they’re done by hand (and forgotten three weeks later), and they only ask “did I show up?” — never “how much original value did I add?” The version Stradiji built closes all three. We build a Claude Cowork scheduled task that runs your brand questions automatically every week across ChatGPT, Perplexity, gemini.google.com, Google AI Mode, and AI Overviews, benchmarks you against competitors, checks whether you’re being cited (citation = the AI linking to your site in its source list), and on top of that scores your content’s information gain (Chris Long’s rubric, Google Patent US11354342B2). Instead of a hand-run audit, you get a self-running measurement system grounded in GEO principles.
First, let’s be clear about what we’re measuring and why
I keep coming back to this: there are two distinct ways to appear in AI search, and you can’t conflate them. A mention is when the AI names your brand as an option inside the answer. A citation is when the AI links to your site in its source list, using your content as evidence. They’re not the same. A mention is brand awareness; a citation is real traffic. Most brands track neither right now.
The second issue is engine diversity. The Search Console report that launched this week only sees Google’s AI surfaces. But your buyer asks ChatGPT, researches on Perplexity, consults Gemini. Each of these engines may know you differently; you can lead on one and be invisible on another. A brand that measures visibility through a single engine is missing half the picture.
The third is competition. Knowing your own impression count is half-knowledge. The real question is: who does the AI recommend for that same query? You, or your competitor? Which sources is your rival cited from while you’re not? That’s exactly where the audit earns its keep: putting your position side by side with your competitor’s.
These three axes (mention/citation, multi-engine, competition) form the backbone of the audit. Now let’s turn it into a task that runs on its own.
Step 1: Localize your prompts and save them to a file
The task’s input is a prompt list tailored to your brand. Fill the prompts below in with your own brand name, category, company size, and target audience. Paste them into a prompts.md file and drop it in a folder Cowork can reach. The set I use for Stradiji has four groups.
Visibility (does the AI know your brand?):
- “Who are the best [category] consultants for [country/city] in the [industry] space? List the top options, explain what each is known for, and tell me which sources you’re drawing on.”
- “As a [job title], I’m looking for [category] services for the first time. What should I know before I start, and which firms should I add to my initial list?”
- “We’re struggling with [the specific problem your brand solves]. What are the most effective ways to fix this, and which firms or tools are considered best at it?”
Brand search (do they describe you correctly?):
- “Who is [your brand], what does it do, what is it expert in? Also tell me where you got the information.”
- “What’s the difference between [your brand] and [competitor]? When should each be preferred?”
Citation check (do you show up as a source?):
- “What are the best sources, guides, or people to follow to stay current in [your category]? List specific sources and explain why they’re authoritative.”
- “What does the research say about [a trend or problem in your category]? Cite the most credible studies, reports, or data sources.”
Competitor benchmark (where do you sit in the consensus?):
- “Who are the main players in [your category]? For each, describe what they’re good at, who their ideal customer is, and their market reputation.”
- “In 2026, how would a [job title] evaluate, select, and implement a [category] service? At each stage, which sources, firms, or content do they turn to most?”
That last prompt is the most diagnostic; it maps where you appear and where you go quiet across the whole buyer journey in a single answer. Save it for last.
Step 2: Write the task brief and name the engines it should scan
Open a new chat in Claude Cowork and describe the task. The brief has to be precise, because the task runs while you’re away. Here’s the brief I use:
“Process every prompt in prompts.md in order. For each prompt, assess how my brand appears in AI search: am I mentioned, am I cited as a source, which competitors stand out, and which sources is the AI pulling from? Do this assessment separately for five engines: ChatGPT, Perplexity, gemini.google.com, Google AI Mode, and Google AI Overviews. Run the same prompts on each engine and compare the results.”
The critical move here is naming the five engines explicitly. Cowork can reach these engines through the browser and run the searches; but if you don’t say which ones you want, it’ll guess. Spell out ChatGPT and Perplexity, Gemini (gemini.google.com), Google AI Mode, and the AI Overviews that appear in classic search, one by one, so the picture is complete.
Step 3: Have it research the sources the AI cites too
A detail that multiplies the audit’s value: if the AI isn’t citing you, who is it citing, and why? Add this to the brief: “For each prompt, list the sources the AI references (sites, reports, YouTube videos, Reddit threads, review platforms). Note which of these sources carry my competitor’s content but not mine.”
This produces a direct roadmap for your content strategy. Say that for a certain question the AI is feeding off three YouTube videos and you have no video on that topic. That single observation answers “what content should I produce this week?” Reddit, YouTube, and review platforms (G2, Capterra and the like) are among the sources AI cites most often; have the task scan those specifically.
Step 4: Have it compute the information gain score too (this is where we differ from the simple audit)
Just checking “did I show up, was I cited?” is half the job. The real question is: why should the AI cite you at all? Google makes that call on information gain — for the same query, how much new information does a page add that the others don’t? Chris Long’s Information Gain rubric (Google Patent US11354342B2) measures this across six dimensions: first-party evidence (invoices, screenshots, your own data), firsthand experience (“I did it, saw it, built it”), specificity (named people, exact numbers, real dates), point of view (content with a thesis), the LLM moat (could ChatGPT write this from training data alone?), and information gain (what are you teaching that isn’t in the top three results?).
Add this to the brief: “For every question where my brand isn’t cited, read the top three organic results, list the information those pages have that mine doesn’t. Then score my relevant content on the six dimensions above and tell me which dimension I’m weak on.” The output won’t just say “I’m not showing up”; it’ll say “that page has a real customer case, yours has generic advice, that’s why it’s cited and you’re not.” That’s a diagnosis you can act on directly. This is the core of Stradiji’s GEO approach: AI doesn’t cite commodity content that anyone could write; it cites non-commodity content that only you could write.
Step 5: Define where the output is saved and when it runs
Add the save rule to the brief: “Write the results to AI_Visibility_Audit/Report_YYYY-MM-DD.md. In the report, give each engine its own section, summarize mention/citation status in a table, and collect competitor benchmarks and content gaps under separate headings.” The dated filename means a baseline accumulates every week; lay two weeks’ reports side by side and you can see the trend.
To set the schedule, just type something like “run this task every Monday at 9 AM” into the chat. Cowork offers to save it as a scheduled task, and you confirm. From then on, five engines get scanned every Monday without you lifting a finger, and the report lands in your folder.
Step 6: Trigger the first run manually and read the report
Start the first report by hand and read it carefully. Do the prompts recognize your brand correctly? Does the competitor list make sense? If one engine doesn’t see you at all, that’s a serious visibility gap in that ecosystem. Refine the prompts if needed and update prompts.md; the task runs with the new version next time.
One caveat: AI answers vary. The same prompt can show different sources on different days. So don’t draw firm conclusions from a single report; gather at least three or four weeks of data and read the trend. The task collects the data; you make the call. The real work is reading that data and updating your content and PR strategy — that part is still a human job.
Three things to watch when you set it up
You can get stuck in three places when you actually build this. First, permissions. The task uses the browser; on the first run Cowork asks for permission for Chrome and other tools. The moment you create the task, trigger it once with “Run now” and approve the permissions up front. Otherwise Monday morning the task sits waiting on a permission screen and never runs. Second, don’t write the source list from memory. Don’t set the task up around “AI feeds off Reddit and YouTube” just because you read that somewhere; it varies by vertical. Have the task ask the engines directly each week, “which sources did you base this recommendation on,” and let it bring back the real sources. Third, build the competitor list correctly. I didn’t guess my competitors; I pulled stradiji.com’s organic competitive set from Semrush (Zeo, Webtures, AnalyticaHouse, SEMpeak, Digipeak). Define your own competitors from data, not memory.
What the first audit actually showed: the AI is being gamed
When I set this task up for Stradiji and ran the first audit, the result didn’t surprise me, but it drew a clear picture. There are two separate worlds.
On brand and personal queries (“who is Stradiji,” “who is Mert Erkal,” “what’s their reputation”), we show up, and the sentiment is positive — meaning the AI uses positive, trust-building language when it talks about us. It knows us from LinkedIn and the client testimonials on stradiji.com, and it finds that we’re described as “the doctor for sites losing traffic to technical errors” on a major local tech forum. No problem there.
But on commercial queries like “best SEO agency / GEO agency in [my market],” we don’t show up at all. The striking part is how the agencies that do show up got there. I looked myself: one agency put a different firm at #1 and itself at #2 in its own blog post titled “the best GEO agencies in [city].” The AI takes that list as a source and presents that agency as “among the best.” So the AI is being flat-out gamed by self-promoting ranked lists (listicles — “top 10 agencies” style content). The same phrases repeat across dozens of sites; this is called recommendation poisoning — spreading the same message across the web to get the AI to present a particular brand as the top recommendation. And the AI mistakes that repetition for “consensus.”
This is GEO’s dirtiest spot right now. It looks like it works in the short term. But as Lily Ray has warned repeatedly in 2026, this kind of synthetic citation building and pile of self-promotional listicles is flagged as a spam signal by Google and Microsoft; it carries algorithmic-penalty risk in the long run. Stradiji won’t play that game. Our path is what the AI actually values: original data, real case studies, firsthand experience, and digital PR with genuine news value. Not stuffing lists, but producing proof.
I poured the results into a dashboard
Instead of reading that report every week, I piped the task’s results into a live dashboard. Every Monday the task both writes the dated report and refreshes this panel. The dashboard shows an overall visibility score, the status on each of the five engines, Stradiji and Mert Erkal tracked separately, sentiment and which source it comes from, a competitor benchmark, the weekly trend, and most importantly a matrix of the audited prompts. Prompts where Stradiji never appears get flagged automatically as “content gaps”; those become the priority list for non-commodity content. The panel is bilingual, English and Turkish, with one click to switch.
The dashboard isn’t meant to replace our GeoGenie.ai panel. GeoGenie is the core measurement layer for our enterprise client campaigns; this Cowork dashboard is a lightweight, fast-to-build complementary layer where I can track my own brand and any prompt I want. The two work together.
Why I keep hammering on this
For three years I’ve been saying “measure your AI visibility” in this newsletter. This week we turned that sentence into a system. Because measurement done once doesn’t work; in this new era the winner isn’t whoever produces the most content, it’s whoever tracks visibility most consistently and adapts fastest. At Stradiji we run our GEO campaigns through GeoGenie.ai; these lightweight Cowork automations we build ourselves run right alongside it, and the two complement each other. This is the strongest lever in your hands too: set it up once, let it run every week, and you just read the report and act.
Görüntülenme Sayısı: 17
