90 Days Into an LLM Visibility Audit: What Happens When You Stop

This is Post 5 in a 90-day LLM visibility audit I've been documenting live on my own brand, NextWise Studio. Post 1 covered the baseline. Post 2 covered what I changed and why. Post 3 covered the 30-day results. Post 4 covered the 60-day results. This post covers Day 90.

I need to start with something the data doesn't explain on its own.

Day 90 was supposed to land on July 20. It's landing five weeks late. Between July 1 and now, I didn't publish a single new post, build a single new citation, or send a single outreach pitch. A vacation and a second startup took the time I'd been putting into this experiment.

I'm telling you that upfront because the numbers only make sense with that context. And because what they show turned out to be more useful than if everything had just kept climbing.

What I measured

Same six sources as before, plus one I brought back. Manual prompts across ChatGPT, Perplexity, Claude, and Google AI Overviews. Google Search Console. Ubersuggest's AI Search Visibility tool. Website analytics. AI crawler activity. And, for the first time since Day 30, a formal set of geo-specific prompts.

Six sources. A hundred and five days, not ninety. Here's what each one showed.

What held or improved

Branded search position is essentially perfect now.

"nextwise studio" now averages position 1.03 in Google Search Console. At both Day 30 and Day 60, that number was 4. This is a real, substantial jump.

The naming collision is easing.

Impressions driven by searches for the Romanian NEXTWISE AI Systems Studio made up 60 to 70 percent of my Search Console data at Day 60. That share is now down to roughly 41 percent. Still the biggest source of noise in this measurement, but moving in the right direction.

Google AI Overviews now disambiguates inside the answer itself.

At Day 60, disambiguation showed up as sidebar links. Now it's in the generated text: Google correctly identifies nextwisestudio.com as the US-based marketing consultancy, then adds a note about the separate European entity. That's a meaningfully stronger signal than a link off to the side.

Perplexity moved from mixed to strong.

Correct attribution, and it now pulls Bay Area and Silicon Valley language directly from my site copy. That's the geo-specific site changes showing up somewhere concrete.

The LLM Visibility Audit topic kept compounding.

50 percent visibility and rank 1 at Day 60. Now 100 percent visibility, rank 1.00, on Ubersuggest's AI Search Visibility tool. The one topic most tied to the actual content of this audit is the one that didn't just hold. It grew.

Backlinks climbed every single checkpoint.

18 at baseline. 37 at Day 60. 44 now. This is the one metric that never paused, and I'll come back to why I think that is.

ChatGPT queries referencing my site nearly doubled.

33 at Day 60. 61 now, a 74 percent increase, the largest jump of the entire experiment.

And for the first time in three checkpoints, I appeared in a category prompt.

ChatGPT named NextWise Studio, unprompted, in 2 of the 6 broad national category prompts I've been tracking since Day 30. Both times, I was listed by name alongside real, established competitors, Accenture, Deloitte, PwC, EY, with an accurate, specific description of what I actually do. Neither mention came with a citation link, so I can't confirm the exact source. But the specificity reads more like genuine recognition than coincidence.

Zero appearances across three checkpoints, then two, in one tool. That's the first real movement on the hardest metric in this whole experiment.

What declined

I want to be direct about this part rather than bury it in qualifiers.

Overall brand visibility on Ubersuggest has now dropped at every checkpoint since Day 30.

20 percent, then 17 percent, then 14 percent. At Day 60 I described the drop to 17 as within variance. With a third data point now moving in the same direction, that framing doesn't hold up. This is a trend.

I lost a category topic I'd broken into.

At Day 30, NextWise Studio ranked Number 5 with 50 percent visibility in the "Marketing Consultant AI Transformation" topic. At Day 90, I don't appear in that topic's brand list at all. It's now Omniscient Digital, QuantumBlack, McKinsey, Slalom, and Single Grain.

Share of Voice slipped from 3.1 percent to 2.7 percent.

Site sessions are back to roughly where they started.

Sessions are down 7.7 percent from my original Day 0 baseline, essentially flat. Bounce rate climbed to 92 percent. Average session duration dropped to 33 seconds. Whatever growth trajectory was building through Day 60 didn't continue.

Claude is exactly where it was at Day 60.

I compared its answers to the same two brand prompts, word for word in substance, across both checkpoints. Same wrong LinkedIn profile. Same failure to connect me to founding NextWise Studio. Same lead with the Romanian entity as the "most likely" match. This is despite the schema sameAs update, Crunchbase verification, and Wikidata entry all going live between Day 30 and Day 90. Whatever signal Claude is weighting most heavily, it isn't the one I've been building.

Geo-specific prompts showed nothing.

I rebuilt this test formally for the first time since an undocumented Day 30 spot check, running three location-based prompts across all four tools. Zero appearances, across all twelve results.

And my own "I Audited My Own LLM Visibility" post still isn't being found.

Zero Search Console impressions for the post. Zero citations across all four LLM tools when I ran its exact title as a prompt. Two independent measurement sources landing on the same gap.

The most interesting finding of the entire experiment

I didn't design this checkpoint as a controlled test. The pause wasn't planned. But looking at what held versus what declined, I think it accidentally became the most useful data point in the whole series.

Everything that declined depends on ongoing, active signal. Share of Voice. The category-topic foothold. Overall Ubersuggest visibility. Daily Search Console impressions. These all require fresh citations and continued promotion to sustain, and when that activity stopped for five weeks, they softened.

Everything that held or improved was riding on momentum already in motion before the pause. Backlinks from directory listings already indexed. ChatGPT query volume built on content already published and still getting surfaced. The LLM Visibility Audit topic, tied to structural work completed well before July.

I think this points to something worth naming plainly: traditional SEO behaves like it compounds. A backlink earned in April is still a backlink in August. Domain authority accumulates and decays slowly. LLM visibility, at least the parts of it tied to live retrieval and citation freshness, behaves more like a channel that needs maintenance. Stop feeding it, and it doesn't hold steady. It fades.

If LLM visibility worked the way some "set it and forget it" pitches imply, none of this should have happened. It should have kept compounding regardless of whether I touched it. It didn't.

What I Learned Measuring This 90-Day LLM Visibility Audit

I went into this experiment expecting to measure growth. I ended up measuring what happens without it, which wasn't the plan, but might be the more honest test.

A single before-and-after data point can't fully separate two different explanations for a decline: normal drift in how these tools update and re-index, versus a direct effect of paused activity. What makes me lean toward the second explanation is the shape of the decline. It wasn't uniform. It hit the specific things tied to ongoing citation activity and mostly spared the things tied to durable, already-absorbed signal. Indexing lag alone doesn't usually erode that selectively.

But I want to hold that conclusion loosely until I have a second data point in the other direction.

What comes next

I'm resuming the work now. Reddit and Quora engagement, the still-unsent Search Engine Journal pitch, a case study that's been sitting half finished and not added to sortlist, and a technical fix for 22 pages that lost their meta descriptions somewhere along the way.

And I'm proposing a Day 120 checkpoint, not as an obligation, but as the other half of this finding. If the metrics that declined this round recover once the work resumes, that's a much stronger, more complete proof of the thesis than a single dip with an explanation attached. Work, pause, resume. If visibility tracks that pattern, I'll have shown both directions of the same claim.

I've also put together a full measurement report if you want to see the data in one place:

View the Day 90 Report

I'm still taking on a small number of brands to run this audit alongside my own. If you're curious what your baseline looks like, get in touch.

Day 120 recovery check, date to be determined.