Stop Falling for SEO and GEO Optimization Myths — What I Learned Building and Validating One Service Over a Year

Soulpapamarketing’s branding and marketing story.

First, I need to be clear about something. Soulpapamarketing is not a company that sells SEO or GEO. We don’t take on search engine optimization contracts. What we recommend to and request from brands is one thing: consistently building articles that tell your brand story in line with the topic. That turns out to be the longest-lasting strategy in both search and AI—that’s the conclusion of this article as well.

The reason I’m writing this is one: lately, sales pitches leading with “SEO optimization” and “GEO optimization” have noticeably increased, and a fair number of them have the exact same structure I saw before. Keyword repetition, bulk backlink purchases, traffic abuse, power blog ranking manipulation. They’ve come back with new names. And it’s typically small business owners and mid-size brands without any way to verify these claims who bear the cost.

Words alone make us sound the same. That’s why I created one service directly and validated it for a year without buying anything. Below is that record, and I’m writing down not just what went well but also what went wrong.

SEO and GEO are now where ‘power blog ranking manipulation’ used to be

You probably remember the pattern from the mid-2010s. There was a market for buying top search results, agencies promised rankings, and every algorithm change broke those promises. What remained was flagged accounts and lost money. When they couldn’t directly raise rankings, they inflated visitor numbers with traffic abuse, and those numbers showed up as results in reports. Many advertisers saw site visits increase but sales stay the same back then.

A significant portion of proposal documents now circulating under “SEO optimization” and “GEO optimization” use the same grammar. They guarantee rankings, promise to get you in the top positions, and don’t explain what they actually do. GEO in particular is harder to validate because the concept is so new. The claim “we’ll get you exposed in AI” mostly lacks measurement basis at this point.

I created a service to verify this

Meetmiddle (meetmiddle.co.kr) is a free web service that calculates the midpoint between two starting locations based on distance or time. I launched it in July 2025.

I’ll be honest: I’m not a developer. Meetmiddle was built with what you might call vibe coding, and I don’t even know the details of the code well. I wasn’t trying to prove some sophisticated technology—I just put out a tool for something people genuinely found inconvenient. That one thing: when scheduling a meeting, awkwardly figuring out where the middle is between two people.

And I decided to buy nothing for it. I didn’t buy backlinks, didn’t stuff keywords, didn’t create artificial traffic, didn’t mass-produce articles, and didn’t use ranking services. I wanted to see what would happen over a year in that state. Feel free to try it yourself. Just enter two starting points—no sign-up required.

One year of actual measurement—from 1 click to 670 per month

Google Search Console monthly performance for the entire period. This is pure organic search with no paid traffic, and the data is unedited month-to-month.

It’s not a chart with only upward movement. Clicks dropped from 53 in November to 47 in December, then from 518 in March to 468 in April. Impressions peaked at 24,101 in June and dropped 12% to 21,115 in July. It’s standard practice in the industry to exclude months like these when creating performance materials, but since that practice is exactly what this article critiques, I included all of them.

Which search terms and what ranking positions

Top search terms based on the most recent 4 weeks (7/12–8/9). There were 103 search terms with impressions, of which 59 averaged rankings 1–3, 40 ranked 4–10, and only 4 ranked outside the top 10.

Looking at just the numbers, it seems like good performance. But how these rankings came about is the most important part of this article.

The ranking didn’t go up—the spot was empty

I can’t honestly say this is a result of optimization. To be honest, I moved into a spot nobody else was occupying. The advertising market tells us this. The Naver search ad keyword tool shows not just monthly search volume per search term, but also how many advertisers are actually placing ads on that search term. Money draws advertisers to competitive positions.

‘Find midpoint’ is searched nearly 20,000 times per month but only one advertiser has placed ads. In contrast, ‘marketing agency’ is searched one-twentieth as much but has 10 ads (the maximum). Related search terms like ‘where to meet’, ‘middle location’, and ‘meeting place’ had zero ad impressions.

There’s only one interpretation: people are searching for it, but nobody sees a way to make money, so nobody built it. There was an empty space. When I put a free tool there, it climbed to the top within months. If I’d put that same effort into ‘marketing agency’, nothing would have happened—and our site is proof of that. More on that later, but Soulpapamarketing’s site averages a ranking of 13.8.

Flipping this perspective becomes a practical guideline. Instead of picking high-volume representative keywords, find search terms that are actually searched but have no advertisers. Those spots exist more often than you’d think, and they’re mostly unanswered questions.

What came next wasn’t links or keywords—it was maintenance

What I actually did over a year was mostly tedious maintenance. To summarize:

• The HTML the server sent had zero internal links. The React-rendered screen showed links everywhere, but the initial crawler response contained none. From a search engine’s perspective, the site structure didn’t exist.
• Non-existent URLs were returning 200 OK. Any random string would get a normal response and the home page. This is called a soft 404, and to crawlers it looks like infinitely generated duplicate pages.
• Out of 1,108 sitemap entries, 1,000 were thin auto-generated pages. Over 90 days, those 1,000 pages combined generated 6 clicks and 394 impressions. I removed them from the sitemap and excluded them from indexing. Now there are 127.
• Everything looked fine on screen. There were no visible problems, so manual review kept missing these issues. I switched to automated weekly inspection scripts.

The number of location pages actually appearing in search results grew from 3 in April to 16 in May, 67 in June, and 73 in July. The critical jump was between May and June, resulting from restored internal links and fixed soft 404s. Sitemap cleanup happened in early July, and growth after that was modest—67 to 73. To avoid inflating my own work’s importance, I’m breaking this down accurately.

But the real discovery was something else entirely

Semrush’s June 2025 study, updated in July, titled AI Search’s Impact on SEO Traffic, converted 500+ digital marketing and SEO topics into search queries and prompts, then analyzed Google AI Overviews, Google AI Mode, ChatGPT, Claude, and Perplexity. A finding directly relevant to this article emerges:

About 90% of web pages cited by ChatGPT ranked outside the top 20 of traditional search for the related query. Perplexity and Google’s LLM also frequently cite lower-ranking pages. The study gives three reasons: First, there’s vastly more content beyond rank 20. Second, AI search is designed to surface information fragments, not full pages, so it focuses on individual paragraph relevance and quality over perceived page quality. Third, LLMs pick the best answer for this user, not the best average result for typical users.

The rest of the findings point the same direction.

• For digital marketing and SEO topics, AI search inbound traffic is projected to exceed traditional search traffic by early 2028. It could happen faster if Google’s default experience becomes AI Mode.
• One AI search visitor is worth 4.4 times a traditional organic search visitor. By conversion rate. Users coming from AI answers have already done their comparison.
• 50% of links in ChatGPT responses point to business and service websites. By individual domain, Quora and Reddit are cited most, but the overall distribution shows half go to companies’ own sites.
• One of the study’s practical recommendations: ‘Many AI crawlers don’t execute JavaScript, so make your site crawlable.’ The zero internal links problem I mentioned earlier applies exactly here. Google does deferred rendering, but AI crawlers can’t read that.

In summary: the money spent on buying top positions isn’t targeting what AI search uses as its main criteria anyway. Pages ranking outside 20 still get cited, and those cited visitors are worth 4.4 times more.

I tested it directly—and half our results fell short too

If I only cite research, that’s just another opinion. So on August 12, 2026, I logged into Perplexity unauthenticated and entered nineteen questions directly, checking the responses. What matters is recommendation questions without the brand name. It’s obvious when you ask by your own name, but potential customers ask without knowing us.

1) When asking by name

The second row is what I need to emphasize. Soulpapamarketing’s Google performance during the same period was 55 clicks per month, average ranking 13.8. By top-ranking standards, that’s nearly a failure. Yet AI search accurately described what this company does. It used a 13.8-ranking site as its source. This breaks the assumption that “if you’re not highly ranked, AI won’t mention you.”

2) When requesting recommendations without the name

Meetmiddle made recommendation lists in three out of five tries. Soulpapamarketing failed all four times. The company writing this article didn’t appear once in recommendation queries about our own industry. It’s unfavorable data, but it’s the most useful for this piece.

The two times Meetmiddle fell out also had clear reasons. When asked for an ‘app’, the web service was filtered out, and when asked for ‘how-to’ rather than services, no service appeared. The answer changes based on question format alone.

3) Why does one work and the other doesn’t?

Both sites’ technical conditions are similar. Soulpapamarketing actually has far more content. The split came from two factors.

First, category breadth. ‘Finding midpoint’ is narrow, service identity ends in one sentence, and domestic competitors number in the dozens. ‘Marketing agency’ has thousands of candidates. When AI builds a list, a narrow candidate pool leaves room for individual businesses; a broad one doesn’t.

Second, and this is decisive: when AI builds recommendation lists, nearly all its sources are comparison and ranking articles written by third parties. The basis for agency recommendation answers was Brunch posts and comparison blogs. This mirrors Semrush’s finding that Quora and Reddit are the most-cited domains in AI Overviews. No matter how well you write on your own site, if there’s no third-party document putting you on a list, you won’t make list-format answers.

4) So when does Soulpapamarketing show up?

Failing all four times on industry recommendations naturally leads to the next question: does that mean we never appear without our brand name? So I changed the angle and added eight more. This time, not recommendations, but narrow practical topics where we’ve actually published articles—converted into questions without the brand name.

One out of eight. As a percentage, that’s not impressive. But which question it was matters. ‘Why does splitting ad sets hurt you in VCG bidding structures?’ is barely searched, and documents framing this angle specifically are rare in Korean. And that answer included the phrase ‘VCG bidding structure and APS ROAS-based practical interpretation.’ Our article’s framework was directly reflected.

Looking at the seven unmentioned questions, the pattern splits. Terms like ‘hero branding’, ‘cross-sell’, and ‘safe zone specs’ are either generic or already have established answers—no room for us. ‘APS ROAS’ came back with shallow answers, which actually means nobody’s documented it properly yet.

And right here, the next abuse market opens

Reading that conclusion, one thought probably jumped out: “well, just get my name in third-party list articles, then.” Right. And someone’s already started offering to do exactly that for you.

This structure is precisely what power blog ranking used to be. Once people realized top search results were mostly blog reviews, a market emerged for mass-producing reviews. Then cafe spam, then backlinks. Now it’s “we’ll get you cited in comparison articles AI reads”—packaged as GEO optimization.

The results will match too. Hollow mentions work at first, then get filtered as models update, and all that’s left is wasted money. AI answers also correct slower than search results. Wrong information gets embedded longer.

The real way third-party coverage happens has always been one thing: build something worth mentioning, then wait for empty space. Meetmiddle made recommendation lists not through PR but because that category lacked a quality free tool. Empty space existed and we filled it. Why we’re not in agency recommendation lists is the same reason: that space is already full, and we haven’t produced something that forces the list to be rewritten. This isn’t a thing money moves up—it’s about finding empty space and spending time.

What to look for when filtering SEO and GEO proposals

Things to confirm before spending money:

1. Do they guarantee rankings? If yes, that’s where you stop. It’s an uncontrollable variable.
2. How many people already occupy the search space you’re targeting? Check monthly search volume and ad impression numbers free with the Naver search ad keyword tool. If ad impressions are maxed out, that space is already full.
3. Is the deliverable ‘ranking’ or ‘repairs’? The quote should itemize what’s being fixed. Internal link structure, response codes, sitemaps, structured data—things you can verify before and after.
4. Is bulk backlink purchase included? Pulling massive links from unrelated sites is riskier than rewarding at this point.
5. If they sell ‘AI exposure’, ask how it’s measured. If they can’t explain the measurement method, they have nothing to sell. At minimum, doing what I did—feeding prompts and screenshotting responses—is baseline proof.
6. If they sell ‘securing AI citations’, ask where and what they’re placing. If it’s putting your name in comparison blogs or forum threads, ask if it’s based on actual experience using your service. If not, that’s power blog reskinned.
7. Check your site’s server response directly. Not browser rendering, but the original HTML. Does it have internal links and title tags, or does it return 404 for non-existent addresses? If these two don’t work, nothing else matters.

To get mentioned even without top rankings

The Semrush research and this direct test point to one direction.

First, make it crawlable. AI crawlers often don’t execute JavaScript, so essential content must be in the server response HTML. This isn’t something to buy—it’s something to repair.

Second, answer questions in single paragraphs. AI takes paragraphs, not pages. Paragraphs with data that read complete without context are what get cited.

Third, answer very narrow questions. The research suggests AI tends to cite content tailored to specific use cases and audiences. The only question where we got cited was barely-searched practical terminology. Competing for top positions on high-volume keywords is far less efficient right now than finding what customers actually ask but nobody answers properly. These used to get skipped because search volume looked too low. That’s changed.

Fourth, maintain your own site. Half of ChatGPT response links point to company and service sites. Building on your own domain still beats hosting on other platforms.

Fifth, accept what’s within your control. Getting into recommendation lists needs third-party coverage, which comes when you’ve published something genuinely useful. Soulpapamarketing failing all four agency recommendation questions proves that—we never closed that gap through site optimization.

Why I’m writing this

We don’t sell this as a product. It wasn’t a service built to sell in the first place. But while “SEO optimization” and “GEO optimization” get consumed without verification, I was bothered that the cost falls on those with no way to validate: small business owners and mid-size brands.

So without knowing what I was doing, I built one service and ran it for a year without buying anything. The conclusion is three lines:

• Top rankings aren’t purchased. You find empty space, fix your site, and rankings follow. The search term Meetmiddle ranks for—19,580 monthly searches with one advertiser—was an empty spot.
• At 13.8 ranking, AI still describes us accurately. Not having top rankings is different from not existing.
• Making broad recommendation lists isn’t money’s domain. It takes empty space and time. We haven’t made it yet. Instead, one article citing no-one’s-written-this-angle earned a mention. That’s the empty space to find.

Which is why what we continue recommending to and requesting from brands stays the same. Don’t buy rankings—document what your brand actually experiences. What you learned building the product, questions customers repeat, details others gloss over. Those articles stick in search and AI eventually. Search volume doesn’t matter on unpopular topics anymore—it’s actually an advantage now.

If you’ve received a proposal, try two things before signing: ask AI your brand name, then ask your industry without the name. The difference between the two answers roughly shows whether that proposal addresses something that actually needs solving. Five minutes.

Frequently Asked Questions

<

p style=”font-size:0.85em;color:#888;margin-top:2em;”>Insights from Soulpapa Marketing — Korea’s digital marketing agency.
Original Korean article: https://soulpapa.co.kr/2026/08/12/seo-geo-optimization-check/