I Asked 3 AI Tools The Same Question About My Client. The GEO Results Were Wild.

Context: a real client, a real GEO problem

I build websites and weird interactive stuff for clients. Not "move this button 3px" work, more like "our online presence is invisible and confused" work.

One of my clients is a brick-and-mortar business with multiple locations across the Netherlands. Think local SEO, citations, Google Business Profiles, and a very patchy history of agencies touching their stuff.

We needed a clear picture of their GEO footprint. Where do they actually show up? What is Google seeing? What is the wider web saying? Basically: what is the reality of this brand on the internet, location by location.

I could do the usual drill. Manual Google searches, Google Maps, scraping SERPs, checking directories. I still do that. But I wanted to see how far I could get by treating three AI tools as reconnaissance agents.

So I ran a simple but very opinionated test.

The exact prompt I used

I asked each tool the same thing, with only the client-identifying details stripped for this post:

Client: <redacted brand> is a Dutch company with multiple physical locations
across the Netherlands, focused on <redacted niche>.

Act as a local SEO analyst.

1. Map out all locations you can reliably infer, including city, address (if available),
   and any Google Business Profile or Maps entries.
2. Identify inconsistent NAP (name, address, phone) information and duplicate entries.
3. List the main citation sources and directories currently mentioning the brand.
4. Highlight obvious GEO / local SEO risks and opportunities.

Only use information you can actually see or strongly infer from current public data.
Cite sources for each claim.

I ran this in:

• ChatGPT (GPT-4, browsing on)
• Perplexity Pro
• Claude 3 Opus (with web access enabled)

Same prompt. Same client. Same day.

What Perplexity did: the aggressive scout

Perplexity went first, because it just feels built for this kind of thing. It acts like a research intern that drinks three espressos and opens 40 tabs.

How it approached the query

Perplexity immediately blasted a list of sources in the right sidebar. I saw:

• Official client site (good)
• Google Maps place pages
• Chamber of Commerce style listings
• Industry-specific directories I had not even heard of
• Old press releases and a random PDF from a partner site

The answer came structured as a kind of report. Locations on top, then NAP consistency, then citations, then risks and opportunities.

What it got right

Three things impressed me.

• Location coverage. It found almost every active location, including one that was missing from the main navigation of the site but still lived in a stray landing page.
• Concrete citations. It listed actual directory URLs with clear labels. Things like: "Found on <directory> with phone number X, which does not match website phone number Y."
• Direct quoting. It pulled snippets from pages so I could quickly sanity-check. Not perfect, but close enough to reality that I could verify in seconds.

This is exactly what I want from an AI doing GEO work. Aggressive crawling. Lots of links. Structured enough that I can spot patterns.

Where it messed up

Perplexity has a "confident but a bit sloppy" streak.

• It merged a closed location with the newer headquarters because both had similar names in two different directories.
• It inferred a "service area" from a generic marketing paragraph that mentioned multiple cities. Those cities had nothing to do with actual physical locations.
• It sometimes treated outdated third-party content as present-day reality, especially PDFs and old event pages.

The citations were useful, but I had to keep asking "what is the date of this info?" and "is this page still live?" It rarely volunteered freshness on its own.

Perplexity summary: the fast mapper

If I want a first-pass GEO map with clickable evidence, Perplexity is my default now. It is messy, but it finds stuff I would miss manually or only hit after an extra hour of digging.

I would not trust it to produce a final list of locations or NAP data. I use it as reconnaissance, not as an authority.

What ChatGPT did: the structured consultant

Then I went to ChatGPT with browsing turned on. I treat GPT-4 as my structured thinker. It is very good at taking messy data and turning it into a coherent model.

How it approached the query

ChatGPT took longer to respond, and the browsing steps were more granular.

It basically did this:

• Hit the official site first, found a "locations" or "contact" page.
• Visited individual location pages and tried to pull address and phone numbers.
• Checked Google Maps for matching entries.
• Looked at a handful of major directories.

The response read like a consultant's deck condensed into text. It grouped everything by location with subheadings and tables.

What it got right

Where ChatGPT excelled was consistency and restraint.

• Clean NAP comparison. It literally made a table: "Website NAP" vs "Google Maps NAP" vs "Directory NAP" for each location. Very few hallucinated details.
• Clear caveats. It would say "I cannot confirm additional locations beyond those listed on the official site and major directories" instead of guessing.
• Helpful strategy layer. Besides raw data, it added context like "This location appears overrepresented in citations compared to others, which might skew authority."

I like this behavior. For client work, I prefer an AI that says "I don't know" over one that creatively fills gaps.

Where it missed

ChatGPT felt very "serious consultant with limited time." That cut both ways.

• It missed the weird long-tail citations that Perplexity dug up, especially niche local directories.
• It did not touch some older subdomains and microsites that still leaked NAP data into the wild.
• It almost ignored social platforms that indirectly influence local trust signals, like a dormant Facebook page with an ancient address.

Also, the browsing tool sometimes timed out on clunky local directory pages. Instead of retrying creatively, it just skipped those sources.

ChatGPT summary: the systems thinker

ChatGPT gave me the cleanest overview. If I want a client-facing summary that is mostly accurate and well framed, this is the model I trust the most.

For raw GEO discovery though, it felt cautious and a bit blind to the long tail.

What Claude did: context-rich, data-light

Claude is my favorite "talk through a problem" model. I use it like a co-founder who reads everything and then gives me a thoughtful essay. For this test though, I needed it to behave like a data hound.

How it approached the query

Claude tried to be very transparent about what it was doing.

It clearly separated:

• "Information from the official website"
• "Information from external sources"
• "Inferences and assumptions"

It browsed the site, then Google Maps, then a couple of bigger directories. Then it spent a lot of time explaining how it evaluated NAP consistency and what an ideal local SEO setup should look like.

What it got right

Claude was easily the best at explaining risk.

• Subtle inconsistencies. It picked up that one location always had the brand name plus a qualifier, while another location was frequently listed as just the brand name alone. That is the kind of thing that causes messy Maps merges.
• Historical drift. It tried to infer timeline. For example, "This older directory entry likely predates the rebrand in YEAR, since it uses the former brand name and logo." Pretty good.
• Process advice. The suggestions were not generic. It said things like "Standardize your naming convention as: <Brand> <City> and push that into GMB, website title tags for location pages, and top 10 directories."

From a strategy perspective, it actually felt closest to how I think.

Where it fell short

Part of me suspects Claude is allergic to ugly local directory HTML. It just did not dig as deep as Perplexity.

• Location coverage was decent but not exhaustive. It missed the same oddball landing page Perplexity surfaced.
• It rarely gave direct URLs for specific citations. Often it just said "Listed in several local business directories" without naming them all.
• It spent too many tokens gently disclaiming limitations instead of showing more raw evidence.

So I ended up with a great conceptual overview sitting on top of a relatively thin evidence layer.

Claude summary: the strategist with soft hands

If I already know the data and I want to stress-test the plan, Claude is perfect.

If I rely on it to surface all the weird GEO data, I will miss stuff.

What each one cited, side by side

Here is a simplified breakdown of what sources showed up uniquely per tool for this client.

• All three found: official site, primary Google Maps entries, a few big directories.
• Perplexity unique finds:
  • Obscure industry-specific local directories with outdated phone numbers.
  • An old PDF brochure on a partner site that still had a deprecated address.
  • A sub-sub-page for a temporarily closed location that the main site no longer linked to.
• ChatGPT unique finds:
  • A specific Google Maps duplicate for one location that Perplexity glossed over.
  • One regional business index that it tried to parse more thoroughly, including category tags.
• Claude unique finds:
  • Inconsistencies in how the brand described itself across About pages and directory blurbs.
  • Hints of an older brand identity hiding behind some live content.

So Perplexity won on breadth. ChatGPT won on structured mapping. Claude won on interpretation.

What this means for how I now work with AI

I used to pick a single "main" AI tool and stick with it. This GEO test killed that idea for client-facing work.

Here is the workflow I use now for local SEO and location-heavy clients.

Step 1: reconnaissance with Perplexity

I start with the rough prompt you saw earlier and let Perplexity loose.

Then I:

• Export or copy the answer into a scratchpad.
• Open every suspicious-looking citation URL in a real browser.
• Mark each as: "current and correct", "current but wrong", "outdated", or "duplicate."

This part still needs a human. AI is not great at understanding "this is technically live, but practically obsolete."

Step 2: structured NAP map with ChatGPT

Once I have a rough list of sources and locations, I feed that into ChatGPT.

Example prompt I actually used on a follow-up client:

Here is a list of known locations and citation URLs for <brand>.
For each location, build a table that compares NAP info across:
- Official website
- Google Maps
- Each citation I listed

Highlight any inconsistencies and suggest one canonical NAP version
per location.

GPT-4 is very good at this reconciliation step. The output is almost a ready-made client report.

Step 3: strategy stress-test with Claude

Then I take the cleaned canonical NAP, plus the messy differences, and ask Claude things like:

Given this situation for <brand> (pasted summary), pressure-test this plan:
1. Standardize naming convention as ...
2. Update Google Business Profiles first.
3. Then fix top 20 citations.
4. Then handle long-tail directories as a batch.

What am I underestimating? Where could this go wrong?

Claude will usually point out edge cases like "This location has very few citations, which might be fine if it is deliberately low priority" or "Consider whether the rebrand created trademark issues in older listings."

This is where its "lawyer who reads everything" energy is valuable.

What each tool missed conceptually

Beyond data points, there were blind spots in how each tool thinks about GEO reality.

• Perplexity missed intent. It never asked "Should this location even show up anymore?" It just saw data and assumed "exists."
• ChatGPT missed messiness. It tried to normalize too fast. Real-world citations are ugly. The model tends to retrofit them into a clean schema that hides edge cases.
• Claude missed scale. It had smart things to say, but I could feel it straining once the list of citations grew. It thinks like a strategist, not a crawler.

If you rely on a single model, you inherit its blind spot.

So which AI wins?

None of them. And that is the point.

For this real client, my final GEO picture came from layering all three:

• Perplexity to find the weird hidden stuff.
• ChatGPT to turn chaos into clean tables and a narrative that a client can read.
• Claude to pressure-test the plan and highlight subtle brand and naming issues.

If you forced me to choose one for local SEO work, I would pick ChatGPT with browsing, primarily because it lies the least aggressively and structures the best.

But I think that is the wrong question. The right question is: how cheap is it to ask a second model "What did the other one miss?"

Answer: not very. And for client work where a bad GEO call can nuke foot traffic for a location, that redundancy is worth it.

How I would run this test if you want to replicate it

If you want to try this on your own client or project, here is the minimal version.

• Prepare a single clear prompt that asks for: locations, NAP consistency, citations, risks.
• Run it in all three tools on the same day.
• Skim for unique URLs and unique insights in each answer. Paste them into one master note.
• Use your own browser to verify a random sample of each tool's claims. Do not skip this.
• Then ask a second tool: "Here is what model X found. What did it miss?"

This last step is where the interesting stuff happens. Models are surprisingly good at critiquing each other's blind spots if you give them permission.

I do not treat any of these tools as oracles. I treat them as half-competent junior analysts that work fast, have different personalities, and occasionally hallucinate entire office locations.

Used together, they saved me hours on this GEO audit. Used alone, any one of them would have left me with a dangerously incomplete map.