The Most Common GEO/AEO Issues for B2B Companies

A friend asked me what everybody should be considering for GEO right now.

Most of my video content is built for TikTok and the attention spans that live there. It’s made me pretty good at the 45-second version of most questions. This is not the 45-second version. Every real answer I have to GEO questions runs long, so here’s the whole thing, in the order I look at it.

The order matters. Most of these depend on the ones above them.

Key Takeaways

• Technical accessibility is the baseline; sites must explicitly allow LLM crawlers like GPTBot and ClaudeBot to be indexed.
• AI Overviews favor concise, 40-60 word direct answers placed immediately under question-based headings.
• Off-site signals drive visibility, as over 85% of brand-related AI citations originate from third-party sources like Reddit and Wikipedia.
• Comparison pages only succeed in GEO when they provide checkable, objective data like current pricing and API limits.

Is the blog dead? Nope, keep it, grow it, love it.

The blog is alive, but like a reanimated zombie, I think it has a new job. Luckily that job ain’t eating people – well, not yet anyway.

It used to be a traffic engine. Today it’s the one place your point of view exists in text, in your own words, at length, on land you don’t rent. When an LLM assembles an answer about your category, it needs something of yours to pull. If you starved the blog in 2025 because traffic flattened, you took away the raw material for everything below.

And if you’re asking, “Andrea, do you really need to tell us to still have a blog on the site?” I would answer: “Yes, Charlie, I do. You’d be surprised how many B2B sites I review that have removed it.”

So, step one is making sure you have a mechanism that will allow for structured, robot-read-friendly text on your site. It’s boring, but critical.

What technical issues hurt AI visibility the most?

Four things, and one of them cancels out everything else in this article if you get it wrong.

Crawler access. Check this today. GPTBot, Google-Extended, PerplexityBot, ClaudeBot. Somebody on your team very reasonably added a robots.txt line in 2023 to keep AI from training on your content. Who remembers the fear? “We don’t want AI TRAINING on our precious, precious insights!” Yikes.

That line is now a measured retrieval penalty: Grossman et al. (SIGIR 2026) found sites blocking Google-Extended are significantly less likely to be retrieved by AI Overviews even when the content is otherwise reachable. Also check whether Cloudflare bot-fight mode is doing it for you quietly. I’ve watched a marketing team spend a quarter on GEO work while security had the door shut the whole time.

Page speed. SUPER important, mobile and desktop. Mostly this is still an SEO play, but the LLM bots don’t love waiting either.

JavaScript trust signals. This one makes me a little crazy. Companies love an animated counter. $4.2M saved for our customers! 12,000 installations! It ticks up from zero when the section scrolls into view, and because it’s JavaScript, zero is what the robot sees. Your single best proof point renders as nothing. I’m sorry to say that if you need to trade structural simplicity for “oh wowza!” creativity, it’s just not worth it. Make your receipts (SEOs call this trust signals) plain, boring, static text on the page.

H1s and H2s. The free alerts from Microsoft and Bing Webmaster Tools will flag your H1s and your meta descriptions. They don’t take a stand on your H2s. H2s are where the extraction happens, so they need to read like the question a human would actually type. They also need to exist. I have a lot of clients where we spend weeks and weeks just updating old blogs where no one took the time to ensure the structure existed.

We created a tool for WordPress, Webflow and Hubspot CMS that automates this. Check it out here:

One more that nobody puts on a technical list: slug and redirect hygiene. On one audit, roughly a quarter of a client’s AI citations were pointing at a retired URL. Webpage hygiene is painful, but so, so important.

Should your content answer fast or cover a topic broadly?

The answer is both. Sorry. No one promised GEO would be easy.

Before I get into on-site and off-site recommendations, there’s a lot of talk about how to write for GEO. What kinds of things LLMs either like or don’t. And to me, it’s a fascinating line to walk. What I see is that some “AI discoverability” wants short, to the point answers, while others actually want a robust corpus of detailed technical and highly-opinionated takes that come from earned experience.

Much of this has to do with what model you’re trying to optimize for. Let’s call a spade a spade, a lot of people are getting Slacks from their boss or their CEO directly saying, “I Googled us and I see we have the #1 spot, but we aren’t showing up at the top.” Le sigh. “The top” of Google (for the majority of non-commercial search terms these days, 64.7% of question-form queries according to Xu, Iqbal & Montgomery, 2026) is now taken up by AI Overviews (AIOs).

If you want to optimize Google’s AIOs, answer a question in the first two sentences. The job of AIOs is to resolve one finite question, and therefore all they want to do is go lift the passage that resolves it. Tight, specific, answer-up-front content wins there.

ChatGPT (and the other LLMs) behave differently. They decompose your question into several smaller searches, retrieve against each one, and stitch the results together. That process is building a picture of what your company thinks, across pages. Breadth and consistency of point of view matter more there than any single tight paragraph.

So the practical answer is that you hone in for AI Overviews and open out for everybody else. A tight answer block earns you the AIO placement. A connected body of work on the topic earns you the synthesis inclusion, citation and, hopefully, recommendation. One page can serve both, but you should know which one you’re optimizing for before you write it.

Which queries want a short answer and which want a corpus?

I think about queries in lots of different way, but in this context, the one that’s important is to think in three bands, moving from “one extractable passage wins” to “you need a body of work.”

The fan-out counts below are directional, from running these queries myself. Perplexity shows you its sub-searches, so you can check any row against your own category in an afternoon.

Why the same question behaves differently in your category

Query type sets the starting position. Topic complexity slides it right.

“What is DNS” fans out to almost nothing, because the answer is settled, universal and part of the LLM’s training data for years now. “What is GEO” is the same grammatical query and fans out considerably, because the definition is contested, the acronyms compete, and the model has to reconcile sources that disagree with each other. Same shape of question. Completely different retrieval behavior, entirely because of the state of the topic.

So there are two questions to ask, not one:

1. Does this question have one settled answer? If yes, write the tight extractable version.
2. Is the topic itself contested or immature? If yes, build the corpus regardless of how simple the question looks.

Which means in an emerging category, almost everything lands in Band 3. That’s uncomfortable… and it’s also why “just add FAQ schema” keeps failing for teams selling something new.

If you want the honest version of how much we actually know about the underlying signals, I wrote that up separately: What We Actually Know About LLM Citation Signals, And What We Don’t.

What on-site content actually gets cited?

Content that is precisely on the question, positioned early, with real numbers in it. That’s the finding from 252,000 controlled trials across six LLMs (Vishwakarma, Kumar & Jamidar, 2026): topical relevance and list position were the strongest drivers, explicit data and recent timestamps gave consistent secondary gains, and formatting-only edits did little (although I’ve seen a LOT of competing research here and still believe for the relatively low lift and agentic nature of updates, they should be prioritized).

Here’s what I recommend, in priority order.

1. Put the question in the slug, the H1, and the first two sentences. In one of our citation exports, pages whose slugs mirrored the question wording took about 22% of citation share despite low domain authority. Precision beats authority, and this costs you nothing but discipline.

2. Use SCQA underneath it. Situation, Complication, Question, Answer. It’s how you build an argument a model can follow. It does fight with answer-first, since classic SCQA lands the answer at the end. The way I resolve it: answer first at the paragraph level, then run Situation, Complication, Question, Answer as the support architecture below it. Readers get the answer, while models gets the reasoning.

3. Do the query fan-out exercise. Take the question you care about most. Run it in ChatGPT and Perplexity and look at the sub-searches they actually performed. You can also do this more robustly and systematically with GEO tools (Profound, SemRush, Adobe’s LLMO). Then check whether you have anything answering each one. Almost everybody covers one or two and has nothing for the rest. That gap is invisible if you’re staring at a keyword list, which is why keyword lists keep telling teams they’re done.

4. Load it with data, dates, and specifics. Statistics, quotations, and cited sources boosted generative-engine visibility by up to 40% (Aggarwal et al., KDD 2024). Price info and recent timestamps showed consistent measured gains. Generic advice gets synthesized without you. Your own number forces citation.

5. Add the extractable formats. TL;DR blocks, Q&A, FAQ sections, comparison tables, definition-first glossary entries. Say the answer in the H2, again in the body, again in the FAQ. Redundancy is the whole point. Just remember these are an amplifier on a relevant page, not a fix for an irrelevant one.

6. Fix the E-E-A-T plumbing. Author bylines, actual credentials, visible dates, links out to awards. Missing in nearly every audit I’ve run, and it’s such a shame, because it’s easy.

7. Sort out re-crawl mechanics. Update lastmod, ping IndexNow, resubmit through Search Console and Bing Webmaster Tools. Then clear your cache, because that’s usually why the sitemap date didn’t change.

8. Ungate the research. A gated asset cannot be read by anything. If the play is being cited as the source on your category, the paywall (even if it’s “just” a form) is working against the investment.

Do comparison and alternatives pages work for GEO?

They work when they carry information nobody else can supply. They fail when they’re a brochure with a checkmark grid… that fibs, positions or otherwise distorts the actual truth.

This is worth slowing down on, because “build comparison pages” is standard GEO advice and it’s incomplete enough to waste a quarter.

For any comparison question, the model already has forum threads and review platforms. Those sources own the subjective layer. What real users (hopefully) think about using the tool, whether support actually answers issues in a timely way, whether the migration hurt. You will never out-source a thousand G2 reviewers on how your product feels to own, and you shouldn’t try.

So the question to ask before you write anything is this: what does an LLM need to go get FROM YOU, because you are the only source for it?

That list is real and most companies leave it empty:

• Current pricing and exactly what sits in which tier. Reviews go stale in a quarter.
• The actual integration list, API limits, data residency, security posture. Checkable facts nobody else maintains.
• Real implementation timeline, and the variables that change it.
• What you deliberately do not do, and who should buy something else.
• What changed in the product recently, since reviews are describing a version from 18 months ago.

The test I’d apply: if a model needed this specific fact to answer a comparison question, is your site the only place it exists in checkable form? Publish that. If a Reddit thread already covers it better than you can, skip the page and go work on the thread.

And the reason the BS versions fail is mechanical. They compete on the subjective layer, where the model has five better-sourced options… and they leave the factual layer blank, which was the one thing the model needed a first-party source for.

Naming who you’re wrong for is also the part that gets pulled most often. It reads as a source. Winning every row reads as marketing.

Why does off-site matter more than your own website in GEO?

Because 85.7% of the citations in an answer about your brand point at sites you don’t own.

That number comes from Zatuchin (2026), analyzing 167,551 URL-grounded citations across 128 brands, 12 markets, and 13 languages. Chen et al. (2025) got there experimentally too, describing a systematic and overwhelming bias toward earned media.

Supply is also concentrated. About 80% of brand citations come from roughly 18% of domains, and Wikipedia was the single most-cited domain in 11 of the 12 languages studied. There is a short list of sources the models lean on. You’re on it or you aren’t.

I unpacked all of this, including why a #1 Google ranking doesn’t carry over, in Why Your Brand Doesn’t Show Up in AI Answers.

Here’s how I’d think about the surfaces.

Reddit and forums. In our own citation set, social sources ran about 28.5% of citations, with Reddit the single most-cited source at 15.8%. It’s heavily engine-dependent, which matters for prioritization: on SaaS comparison prompts, Reddit was roughly 24% of ChatGPT’s top-5 citations against 12 to 14% on the other engines.

Review platforms. G2, Capterra, Gartner Peer Insights. This one is conditional, and the condition is your category. Across 16 client audits, review sites showed up about 9 times more often in SaaS comparison answers than in physical and industrial B2B, where they were essentially absent from ChatGPT citations. If you sell software, this is a core investment. If you sell compressed air systems, put the money somewhere the models are actually looking.

Category maturity changes the answer again. Mature categories in our data showed about 34% review-site presence. AI-native categories showed 4% review sites and 36% Reddit. Same broad industry, opposite priority, depending on how long the category has existed. Diagnose before you allocate. There’s more of this in the GEO for SaaS webinar.

Wikipedia and Wikidata. Most-cited domain in 11 of 12 languages. Go look for jobs to manage Wikipedia. Big companies are paying six figures to fill this role.

YouTube. Citation tracks with topical precision, not reach. A 400-subscriber channel answering one narrow question gets pulled while the big brand channel gets ignored. We are an example of this. My short from last year is cited for this search query.

Digital PR, earned media, trade publications. News is a heavy citation category. PR has been doing this work for decades and has never once been measured on it.

LinkedIn, analyst relations, and third-party roundups. You don’t have to write the “best X” list. You have to be in it.

Why doesn’t this off-site work get done?

Partly ownership. Reddit, PR, community, and analyst relations live with three different teams, and whoever owns GEO usually owns none of them. Influence without authority is the actual job description, which is why it stalls.

The less interesting reason is that off-site work is slower and harder to put on a dashboard. On-site work shows movement in six weeks. Earned authority compounds over quarters, and quarters lose to next month’s report every single time.

How should you measure GEO?

Per engine, and never in aggregate.

Grossman et al. measured less than 0.2 average Jaccard similarity between the sources retrieved by Google Search, AI Overviews, and Gemini for the same queries. Three systems, one company, barely agreeing on what to cite. So there is no single “AI visibility” number to put in the deck, and anybody selling you one is smoothing over the thing you needed to know.

Two more measurement realities worth planning around:

Retrieval is unstable by design. The same study found AI Overviews are inconsistent between two runs of the same query and sensitive to minor rephrasing. Your brand can show up Monday and vanish Tuesday. That’s the behavior of the system, not noise in your tool, so give any tracking run four or five days before you conclude anything.

Roughly 70% of AI-sourced traffic arrives with no referrer. The majority of buyers don’t click through on citations; instead they create their short list then still go to Google to validate their findings. So GA4 undercounts you, and an unexplained jump in Direct is often the real signal. Worth wiring up before you get asked for a number.

The way around this is the same dark social. Self-reported attribution touchpoints of the contact us form fill and the sales rep asking, “How did you hear about us?” on the first call. Nothing magic.

What carries over from B2C that might still apply to your B2B use case?

Some of the local search behavior, if you have physical locations.

On a home services audit, AI Overviews fired on only 13% of location-specific keywords versus 71% of non-geo informational keywords, and the Local Pack held 82% of location queries. That’s a “near me” dynamic, so it’s mostly a B2C pattern.

But if you have a service area, a distributor network, or branches, the carryover is useful: location queries are still a traditional local SEO problem (location pages, schema, GBP), and your GEO investment pays off on the informational side. Two playbooks, sorted by query type, which is also the argument in why GEO has to start with the buyer.

If you only do four things this month

1. Check robots.txt and your CDN bot settings. Blocking is a measured penalty and this takes ten minutes.
2. Pick your three highest-value questions and run the fan-out exercise. You’ll find the gap in an hour.
3. Rewrite one page so the question is in the slug, the H1, and the first two sentences, with two of your own numbers in it. Watch what happens (Bing Webmaster will show you this the fastest and it’s free)
4. Look at where your category’s citations are actually coming from, then move a real portion of the investment off-site. The 85.7% is the map.

That’s what I’d tell any B2B team asking me right now. The technical work is real and it’s table stakes. The on-site work is where most teams stop. The off-site work is where the returns are, and it’s the part sitting unowned in almost every org chart I’ve seen.

Pick the one you’ve been avoiding. That’s usually the one paying the most.

Frequently Asked Questions

How do I check if my site is blocking AI crawlers? You should review your robots.txt file for directives that disallow GPTBot, ClaudeBot, or Google-Extended. Even if your standard SEO is healthy, these specific robots must be permitted for your content to surface in LLM responses and AI Overviews.

What is the best word count for a GEO-optimized paragraph? For direct answers meant to be extracted by AI Overviews, aim for 40 to 60 words in the first sentence or two under a heading. This matches the ‘extractable zone’ pattern where engines seek concise, factual responses to definitional or binary queries.

Why does Reddit rank so high in AI citations? LLMs prioritize human-led discussions and community validation, which Reddit provides at scale. According to current research, social sources account for nearly 29% of citations in some categories, making off-site presence as critical as your own website’s content.