# Your LLM SEO strategy needs fewer, stronger pages

> A practical LLM SEO strategy for B2B SaaS: build citable comparison and pricing pages, fix access, measure citations, and stop publishing filler.

An LLM SEO strategy for B2B SaaS should produce fewer pages, each built to settle a real buying question with facts that a retrieval system can quote without repairing your logic. Publishing hundreds of lightly differentiated articles is not a shortcut to citations. It creates a site full of interchangeable claims, unclear ownership, and facts that disagree across URLs.

The pages with the best chance of earning citations are usually close to a decision: a transparent pricing page, a fair comparison, a technical integration page, an original benchmark, a security or compliance explanation, and documentation that names exact constraints. They work because each page answers a bounded question and supplies evidence. The same traits help a buyer, which is why good LLM visibility and good product marketing often point in the same direction.

## Citation is a retrieval outcome, not a writing style

An LLM cites a page when a search or retrieval layer finds that page relevant to a prompt, can access and parse it, and selects it as support for part of the generated answer. The model's training data and its live search sources are different things. Blocking a training crawler does not necessarily block a search crawler, and inclusion in training does not make a page eligible for a current citation. Teams that blur those two systems end up changing the wrong controls.

Google Search Central explains that AI Overviews and AI Mode can issue multiple related searches across subtopics and data sources through query fan-out. That matters for B2B SaaS. A prompt such as "best billing platform for a European marketplace" can produce retrievals about marketplace payouts, VAT support, pricing, regional availability, migration, and alternatives. One broad landing page rarely supports every part. A tight cluster of canonical pages can.

OpenAI's publisher guidance says a public page can appear in ChatGPT search and that OAI-SearchBot must be allowed for content to appear in summaries and snippets. Perplexity makes a similar distinction in its crawler documentation: PerplexityBot surfaces and links websites in search results and is separate from crawling for foundation models. These manuals do not promise a citation. They define eligibility, which is the part you can control.

This gives a useful operating model. Citation visibility has four gates:

1. The crawler can fetch the canonical URL and see the main text.
2. The index understands which entity, problem, and claims the page covers.
3. Retrieval ranks the page for one of the prompt's explicit or fan-out questions.
4. The answer system chooses a passage as support and displays a citation.

Content work mostly improves gates two and three. Technical SEO protects gate one. Nobody outside the answer system controls gate four, so any agency guaranteeing citations is selling confidence it cannot verify.

## Decision pages beat generic educational posts

The most citable B2B SaaS pages carry facts that are both specific and useful at decision time. Generic educational posts often explain a category with the same definitions found on dozens of sites. An answer system gains little by citing the forty-first version. A page becomes useful as a source when removing it would remove a fact, a test, or a defensible point of view from the web.

Prioritize these page types:

- Product and pricing pages with plan boundaries, billing units, included usage, overages, and worked cost examples.
- Comparison and alternative pages that state who should choose each option, use consistent criteria, and admit where the competitor fits better.
- Integration and documentation pages with prerequisites, supported objects, error behavior, limits, and copyable examples.
- Trust pages covering security architecture, data handling, compliance scope, availability definitions, and the evidence behind each claim.
- Original research, migration reports, benchmarks, and calculators whose method is visible enough to inspect.

A category explainer can still earn citations when it introduces a sharp distinction or original artifact. Most do neither. "What is revenue automation?" adds little if it paraphrases existing definitions. "How revenue automation handles a partial refund after a plan change" owns a concrete edge case. The second page gives retrieval a recognizable problem, and it gives a buyer something to test.

Do not turn this list into six thin content programs. Start with the questions that appear in sales calls, procurement reviews, support tickets, implementation plans, and lost-deal notes. Those inputs reveal what buyers cannot resolve from the current site. Search-volume tools tend to understate narrow commercial questions, while those questions often carry the strongest purchase intent.

A useful inventory labels every candidate page by question, buyer stage, fact owner, proof source, and canonical URL. If two pages answer the same question with the same evidence, combine them. If a proposed page has no internal fact owner, it will probably become a summary of other summaries. Do not publish it.

## Comparison pages must help the buyer disqualify you

A credible comparison page lets the wrong buyer reject your product quickly. That sounds uncomfortable because many SaaS teams treat comparison content as a sales ambush. Their table gives their product twelve green checks, gives the competitor six red crosses, and hides the criteria in vague labels such as "flexibility" or "enterprise ready." A buyer distrusts it, and a retrieval system has little precise material to quote.

Choose comparison criteria from the job the buyer needs to complete. For an event-processing product, criteria might include delivery semantics, replay window, regional storage, schema handling, throughput controls, and failure recovery. For a support platform, they might include channel coverage, routing rules, identity model, audit history, and pricing unit. Each criterion needs an observable definition.

State the basis and date of the comparison near the table. Link-free source notes in the prose can name public documentation, direct product testing, or vendor-confirmed information. Mark anything you could not verify as unknown. Never convert "not documented" into "not supported." Those claims are different, and confusing them damages the entire page.

Use this copyable structure for each serious competitor page:

```text
# [Your product] compared with [Alternative] for [specific job]

Short answer: two or three sentences naming the decisive difference.

Who should choose [Your product]
Who should choose [Alternative]

Comparison criteria
| Criterion | How it is measured | Your product | Alternative | Source checked |

Cost scenarios
- Small but usage-heavy customer
- Large customer with light usage

Migration constraints
Known unknowns
Last verified: YYYY-MM-DD
```

The failure this prevents is easy to recognize. Without "how it is measured," writers compare a published maximum with a default limit, or a list price with a negotiated enterprise quote. Without cost scenarios, they announce that one product is cheaper even though the pricing units differ. Without known unknowns, they fill gaps with inference.

Write one page per competitor only when the decision criteria or migration path genuinely changes. Otherwise, a single alternatives page with a decision table may answer the question better. Do not create city, industry, and company-size permutations that repeat the same verdict. Google now describes such fan-out targeting, when produced mainly to manipulate generative answers, as scaled content abuse. It is also miserable for buyers.

## Pricing pages need arithmetic, boundaries, and dates

A pricing page earns trust when a buyer can calculate a plausible bill without opening a sales chat. The plan cards are only the beginning. LLM prompts frequently ask what a product costs for a particular team size or workload. A page that says "contact sales" beside three undefined tiers cannot support that answer. The model may cite an aggregator, an old review, or a forum comment because those sources contain the missing number.

Name the pricing unit in plain language. Is the customer paying per seat, active user, workspace, transaction, token, gigabyte, managed account, or a combination? Define when the unit is counted and whether inactive objects still count. Then state the billing period, minimum commitment, included allowance, overage rule, currency, tax treatment where relevant, and whether annual pricing requires prepayment.

Show at least two worked examples that use real plan rules. One should represent a common account. The other should expose a boundary, such as a usage spike, an extra workspace, or a feature that forces an upgrade. Keep the arithmetic visible:

```text
Monthly total = base plan + billable seats + usage overage
Example = $400 + (18 seats x $25) + (120 units x $0.08)
        = $859.60 per month before tax
```

That example is a format, not permission to invent numbers. Product, finance, and legal owners must approve every input. If enterprise pricing is negotiated, say which variables affect the quote and give a representative scenario only when the business can stand behind it. A range without assumptions creates more confusion than "contact sales."

Keep historical prices out of the canonical pricing page unless customers need them. Put legacy plans on a clearly dated support page and identify who remains eligible. When a price changes, update the visible text, structured data, sales collateral, help center, and third-party profiles in the same release. Conflicting prices force retrieval systems to choose among sources, and they may choose the stale one.

SoftwareApplication or Product markup with Offer data can clarify machine-readable price information when it accurately reflects the page. It cannot rescue hidden or misleading pricing. Google explicitly says structured data must match visible text and that no special schema is required for its AI features. Treat markup as a typed copy of the facts, not a secret message to an LLM.

## Technical access fails before content gets a chance

A beautifully written page cannot earn a live citation if the relevant crawler receives a block page, an empty client-side shell, or a canonical tag pointing elsewhere. Marketing often checks the page in a logged-in browser and declares it accessible. Crawlers see a different route through the CDN, web application firewall, consent layer, and rendering stack.

Start with the public canonical URL in a clean session. It should return a successful status, a stable canonical, indexable directives, and the main answer in HTML. Important facts should not require a click, form submission, region selector, or script that the crawler cannot execute. If the page varies by country, make each version addressable and explain the currency and availability in visible text.

Then test the actual user agents and inspect server logs. A basic check looks like this:

```bash
curl -I -A "OAI-SearchBot" https://example.com/pricing
curl -I -A "PerplexityBot" https://example.com/pricing
curl -s -A "OAI-SearchBot" https://example.com/pricing | head
```

The expected header shape begins with a successful response and the intended content type:

```text
HTTP/2 200
content-type: text/html; charset=utf-8
link: <https://example.com/pricing>; rel="canonical"
```

A 200 response alone proves little. Some bot defenses return a challenge page with status 200. Search the body for the page's H1 and a distinctive pricing sentence, then compare it with what an ordinary visitor receives. Review logs for the published crawler IP ranges before allowlisting by user-agent string alone, because strings can be spoofed.

Check robots.txt separately for Googlebot, Bingbot, OAI-SearchBot, and PerplexityBot. Decide deliberately which search crawlers and training crawlers you allow. Do not paste a viral robots file without understanding those distinctions. Preview controls such as noindex, nosnippet, and max-snippet can also limit what an answer system can display even when crawling succeeds.

JavaScript is not automatically fatal, but hiding the central answer behind delayed rendering adds failure modes you do not need. Server-render the title, headings, tables, prices, definitions, and source notes. Use scripts for interaction around that content. The HTML should still state the answer when the calculator widget fails.

## Clear passages make claims easier to reuse accurately

A citable passage answers one narrow question, names the subject, qualifies the claim, and keeps its evidence nearby. This is less about feeding chunks to an LLM than about writing prose that survives extraction. If a paragraph begins with "It" and depends on three earlier sections to reveal what "it" means, a retrieved excerpt can become ambiguous. Repeat the entity name when clarity requires it.

Put the direct answer in the first sentence after an H2. Follow it with scope, evidence, exceptions, and an example. A comparison section should name both products. A pricing section should include the currency and billing period. A security claim should distinguish the certified entity, covered service, and current scope. Tight qualification makes a claim more useful, not less.

Tables work when readers need repeated comparisons across exact fields. They fail when a writer squeezes paragraphs into cells or uses icons without text. Give every column one meaning, keep units in labels or values, and write "unknown" when evidence is absent. Repeat the decisive result in prose because some retrieval systems may flatten a table badly.

Entity consistency matters across the site. Use one current product name, one company description, and one canonical description of each plan. Explain former names where necessary instead of silently mixing them. Keep author pages specific about experience and responsibility. A first-person claim from a named operator is easier to assess than a floating assertion attributed to "the team."

Evidence should sit close to the claim it supports. For internal evidence, name the method: production logs over a stated interval, a test setup with a published configuration, or invoices sampled under defined rules. For external evidence, name the manual, standard, or vendor document and explain what it establishes. A pile of source names at the bottom forces both buyers and retrieval systems to reconstruct the argument.

Freshness labels deserve the same precision. "Updated recently" says nothing. Put a verification date beside volatile comparisons, prices, integration limits, and regulatory claims. Do not change a page's date because someone fixed punctuation. Record what material fact changed, even if the note is one sentence.

## Stop writing pages that contain no owned information

Stop publishing glossary entries, trend reactions, and long-tail variants when your team adds no experience, data, or decision support. These pages are cheap to generate because they require no access to the company. That is exactly why every competitor can generate them too. Volume does not create authority when the pages are substitutable.

The popular recommendation to create a page for every question in a prompt fan-out is wrong. It sounds scientific because query expansion is real, but it confuses retrieval behavior with a publishing plan. Google Search Central's generative AI guidance warns against producing separate pages for query variations mainly to manipulate rankings or generative responses. One complete page can answer a family of related questions without splitting authority across near duplicates.

Also stop writing:

- Competitor pages whose research consists of reading the competitor's homepage.
- Definition posts that never connect the term to a product decision or operational consequence.
- Fake benchmarks with undisclosed data sets, configurations, exclusions, or sponsors.
- Annual "trends" posts that change the year and preserve every recommendation.
- Integration pages for combinations nobody has installed, tested, or agreed to support.

Deleting or consolidating weak pages is not an automatic ranking tactic. It is an editorial correction. Redirect a removed URL only when another page truly satisfies the same intent. Otherwise, return a clear removal status and fix internal links. A mass redirect to the home page hides the cleanup from your own reports and sends users somewhere useless.

AI-assisted drafting can still help a subject expert outline, challenge gaps, normalize tables, or turn test notes into prose. The owner of the facts must review the output. If nobody inside the company can tell whether a paragraph is true, the company should not publish it under its name.

## Measure citations as a diagnostic, not a vanity score

Citation measurement should connect a cited URL and prompt family to qualified visits, sales conversations, and corrected content. A single site-wide "AI visibility score" hides the information a team needs. It can rise because an irrelevant definition page appeared in many low-value answers while decision pages remained absent.

Build a stable prompt set from real buyer language. Include category discovery, alternatives, direct comparisons, price scenarios, migration risk, integration fit, security review, and one or two prompts where your product should not be recommended. Run prompts in clean sessions and the relevant market or language. Record the answer surface, date, model or mode when visible, cited URLs, your citation presence, factual accuracy, and answer position only if the interface makes it unambiguous.

Sampling has limits. Answers change, personalization differs, and interfaces run experiments. Treat each run as an observation, not a rank report. Weekly or monthly checks on a fixed prompt set reveal useful changes without pretending that a daily sample represents the whole system. Store screenshots or raw exports where terms permit so a later reviewer can inspect what the answer actually said.

Bing Webmaster Tools now reports total citations, cited pages, and samples of grounding-query phrases across supported Microsoft AI experiences. Bing explicitly says those aggregate metrics do not indicate a page's rank, authority, or role in a specific answer. That qualification is useful: combine platform data with referral logs and manual answer review instead of turning citation count into a new version of impressions.

Segment referrals from answer engines, but do not judge the channel only by last-click volume. A buyer may read a cited comparison, return directly, and book later. Ask new opportunities how they researched the category, preserve referrer data where available, and compare assisted conversions for the pages you deliberately improved. The commercial question is whether accurate visibility moves the right buyer forward.

When a citation is wrong, identify the failure class. A stale fact calls for synchronized updates and recrawling. An absent page may have an access or retrieval problem. A competitor citation may simply contain better evidence. A hallucinated claim with no supporting URL is not an SEO defect you can repair with another paragraph.

### Separate mentions, citations, visits, and outcomes

Measure four different events instead of calling all of them visibility. A brand mention means the answer names the company or product. A citation means the interface attaches one of your URLs as support. A referral means a person follows that source to the site. A commercial outcome means that visit contributes to a signup, qualified conversation, pipeline movement, or retained customer. These events can move independently. A model can mention a famous brand without citing its site, or cite a detailed documentation page without naming the vendor in the answer text.

Use URL-level cohorts to keep the diagnosis honest. Put pricing, comparisons, documentation, research, and educational posts in separate groups. For each group, track eligible indexed pages, cited pages, citations, referred sessions, engaged visits, and commercial outcomes. The denominators matter. Ten cited documentation pages out of twelve tell a different story from ten cited blog posts out of two thousand. Do not compare the raw totals as though the content programs have equal scope.

Before changing a page, save its visible text, crawl response, structured data, prompt observations, referral baseline, and conversion behavior. Write down the defect you expect the change to fix. A pricing rewrite might target missing cost scenarios; a comparison update might target an unsupported feature claim; a technical fix might target a crawler challenge. Make one coherent release, record the date, request recrawling where the search engine supports it, and allow enough time for discovery.

The result still will not prove simple causation because answer systems and competitors change at the same time. It can provide a strong operational signal. If the updated URL begins appearing for the intended prompt family, cites the corrected passage, attracts relevant visits, and improves sales-call accuracy, keep the pattern. If citations rise while buyers bounce or arrive with the wrong expectation, the page may be easy to extract but commercially misleading. Repair the answer instead of celebrating the graph.

A useful citation review reads the cited passage in context. Check whether the answer preserved the subject, units, geography, plan, and date. Check whether the source URL is canonical and whether the answer attributed your original evidence to you. A citation that supports the wrong claim is a quality incident, even if a dashboard counts it as success. Send factual errors back to the page owner and preserve the observation for the next run.

Competitor citations also give better editorial input than a generic gap score. Record what the competing page supplied that yours did not: a direct number, a current date, a clearer definition, primary evidence, or a better-scoped answer. Sometimes the competing citation is correct and deserves to win. Copying its headings will not close the gap; obtaining and publishing the missing fact might.

Set review thresholds around business risk. Pricing, security, privacy, and availability claims deserve immediate attention when an answer is wrong. A missing citation for a broad educational prompt can wait. This prevents the team from spending a week chasing a fluctuating mention while a stale pricing page teaches prospects the wrong billing model.

## Build the source library before expanding the blog

A durable LLM SEO program begins with a small source library owned by product, engineering, finance, security, and customer-facing teams. Marketing turns those facts into clear pages, but it should not invent the facts. Give every volatile claim an owner, a verification date, and a defined update trigger.

For most B2B SaaS sites, the first release should cover the canonical product explanation, pricing mechanics, two or three comparisons tied to frequent deals, core integrations, security and data handling, migration constraints, and one original operational artifact. This may be fewer than twenty pages. If those pages are vague, publishing another hundred articles will not fix the site.

Review the library against sales calls and support escalations. When the same unresolved question appears twice, decide whether an existing page needs a better passage or a new canonical page owns a genuinely different intent. That rule keeps the site compact without forcing unrelated questions into one enormous guide.

At oleg.is, I would treat this source-library review as part of the same operating audit used to find wasted engineering and marketing effort: find duplicated work, name the owner, and keep only what can be verified. The work is less glamorous than generating a content calendar, but it leaves the company with assets that sales, support, buyers, and answer engines can all use.

The first action is concrete. Pick the ten buyer questions most likely to affect a deal, map each to the best current URL, and mark the answer as verified, incomplete, conflicting, or absent. Repair the conflicts before drafting anything new. A site becomes citeable when its own team can point to one page and say, with evidence, "that is our answer."