AI Search Citation Case Study in Pakistan

The Client

A Karachi-based B2B SaaS company that builds HR, payroll, and workforce-management software for Pakistani small and mid-size businesses. The platform handles monthly payroll runs, leave and attendance, statutory compliance work (EOBI enrolment, social-security contributions, income-tax withholding statements), and an employee self-service portal. At the point of engagement the business had roughly 2,800 paying SME customers across Sindh, Punjab, and Khyber Pakhtunkhwa, an average revenue per account between PKR 8,000 and PKR 12,000 a month, and a board-level target to double the customer base inside eighteen months.

Go-to-market was a mix of founder-led sales, a modest paid program of about PKR 1.4M a month across Google Search, LinkedIn, and Meta, and a content engine publishing two long-form articles each week. The team had invested seriously in classic SEO and had reached page one for a cluster of branded and adjacent terms. What they had not done was think clearly about where a Pakistani SME founder or HR manager actually starts research in 2026: by typing a question into an AI assistant.

The engagement began after the head of growth ran a simple test. She asked three AI assistants — ChatGPT, Perplexity, and Gemini — “best payroll software for small business in Pakistan.” The company was not mentioned in any of the twelve answers she collected. Two Pakistani competitors appeared consistently. The rest of the answers were generic international tools that do not handle Pakistani tax and compliance logic at all. That gap, not a rankings dip, was the real growth blocker.

The Problem: Invisible Where the Buyer Now Starts

Four issues were quietly capping the company’s pipeline:

• Near-zero citation share. Across a tracked set of 60 priority commercial queries — things like “payroll software for SMEs in Pakistan,” “EOBI compliance software,” “HR software Karachi,” and “best payroll system for manufacturers in Lahore” — the brand was cited in only 8% of AI answers. Competitors with weaker products but more answer-friendly content appeared two to three times more often.
• Content built for keywords, not extraction. Articles were long, well-researched, and optimised for search intent, but they were written as essays. Summarisation models could not reliably extract a single clean answer to “What does it cost?” or “Does it handle EOBI?” because the relevant facts were buried in prose and spread across sections.
• AI crawlers partially blocked. An inherited web-application firewall and an overly strict robots.txt were rate-limiting and in some cases outright blocking the major AI crawlers. The site was being indexed for classic search but under-crawled by the models generating the answers buyers actually read.
• No measurement. Demo requests that arrived because a buyer had seen the brand in an AI answer showed up in analytics as direct or branded-search traffic. The team had no way to prove that AI visibility was worth investing in, so it never got budget.

This is the core pattern we see across SaaS marketing in Pakistan: classic SEO looks healthy, paid is steady, and the fastest-growing source of commercial intent — AI answers — is invisible and unmeasured.

There was also a measurement trap underneath the visibility problem. Because the team had no citation tracking, any demand generated by AI answers was landing in the “direct” or branded-search bucket and being credited to other channels. AI visibility looked like a cost centre with no return, when in reality it was already producing demos the company could not see. Fixing the visibility gap and fixing the attribution gap were the same project — you cannot claim credit for what you do not surface, and you cannot justify investment in what you cannot claim credit for.

Phase 1 — AI Visibility Audit and Crawl Access (Weeks 1–2)

The first two weeks were diagnostic. You cannot fix what you have not measured, and the company had never measured citation share systematically.

Building the citation baseline. We assembled a tracked query set of 60 priority commercial queries, stratified into three tiers: 20 high-intent commercial queries (“payroll software Pakistan price”), 20 problem queries (“how to calculate EOBI contribution”), and 20 comparison queries (“payroll software vs Excel for SMEs”). Each query was run weekly against five answer surfaces — ChatGPT, Perplexity, Gemini, Google AI Overviews, and Microsoft Copilot — and every answer was scored for whether the brand was cited, mentioned without citation, or absent. This produced the 8% baseline and gave us a defensible scoreboard for the rest of the engagement.

The tiering mattered as much as the tracking. Commercial queries are where citations turn directly into demos, so they carried the most weight in the scorecard. Problem queries are where the brand could earn authority citations that bleed into commercial answers later. Comparison queries are list-shaped by nature, which made them the fastest path to being named alongside competitors even before the brand could win a direct-answer citation. Scoring each tier separately let the team see which kind of visibility was moving and which was stuck, rather than collapsing everything into a single number that hid the real story.

Crawl-access audit. Server logs showed the major AI crawlers were hitting the site but a large share of requests were returning 403 or 429 status codes from the WAF. We reviewed the full generative engine optimization and AI discoverability checklist and produced an access fix list.

Access issue	Before	After (Week 2)
AI crawler requests blocked/rate-limited	~41% of hits	Under 4%
robots.txt explicitly disallowing crawlers	2 AI user-agents	0 (explicit allow)
Key commercial pages in AI index	12 of 60 queries’ target pages	47 of 60
Structured data describing the product	None	Draft schema mapped

Entity and content inventory. We catalogued every commercial page and tagged whether it contained a clean, extractable answer to the question a buyer would actually ask. Only 9 of 60 target pages passed. The remaining 51 needed restructuring — not necessarily rewriting from scratch, but reorganising so the answer came first and the supporting detail came after.

The pass rate exposed a deeper habit in how the team wrote. Their content was accurate and thorough, but it was structured for a human reader who scrolls — assumptions, narrative, payoff. A summarisation model does not scroll patiently; it extracts. Pages failed the audit not because the information was missing but because the single sentence that answered the buyer’s question was embedded in the fourth paragraph of a section headed with a clever headline. The fix was structural, not creative, which is why Phase 2 moved as fast as it did.

By the end of Phase 1 the company had a measurement system, open crawl access, and a prioritised list of 51 pages to restructure. No content had been published yet, but the foundation was in place.

Phase 2 — Content and Entity Restructuring (Weeks 3–6)

With access fixed and measurement running, we restructured how the site communicated facts. The principle was simple: an AI summarisation model should be able to answer “What is this, what does it cost, and who is it for?” by reading the first screen of any commercial page.

Answer-first page architecture. Each of the 51 prioritised pages was rebuilt around a tight answer block at the top — a 40-to-60-word direct answer to the page’s core question, followed by a short feature list, a pricing range, and a “who this fits” line. The long-form detail moved below, structured with clear H2 questions. This served two audiences simultaneously: human buyers who want a fast scan, and models that extract the lead answer for citation.

Pricing and feature schema. We implemented product, Offer, SoftwareApplication, and FAQPage schema across commercial pages, plus an organisation-level entity definition describing the company, its product categories, and its relationships (integrations, supported compliance frameworks, customer verticals). This is the connective tissue that helps a model place the brand inside an answer rather than treat it as an undifferentiated string of text. The pattern mirrors the work in our SEO for AI Overviews service.

Two schema details had outsized effects. First, the FAQPage blocks were rewritten so each answer was a self-contained 30-to-50-word statement that matched, almost verbatim, the phrasing a buyer would use to ask the question. Models reach for FAQ answers because they are pre-formed, citable statements; vague or hedged answers get skipped. Second, the SoftwareApplication markup carried explicit applicationCategory, operatingSystem, and offers fields, which gave the models structured facts about what the product is and what it costs — the two things buyers most often ask about and the two things the old pages buried deepest.

Comparison and list-inclusion content. AI answers cite two things disproportionately: direct answers to the question asked, and list-style comparison content that already names the options. We built out ten structured comparison pages (“payroll software for manufacturers,” “HR software for clinics,” “payroll tools that handle EOBI”) formatted as clear, neutral comparisons — including competitors — because answer engines reward comprehensive, unbiased source pages over self-serving ones.

Phase 2 checkpoint (Week 6):

• 51 commercial pages restructured to answer-first architecture
• Product, pricing, and FAQ schema live on all commercial pages
• 10 comparison pages published
• Citation share moved from 8% to 19% as restructured pages were reprocessed

The jump to 19% was the first proof point. It came entirely from existing content being made extractable — no new domain authority, no link building, just clarity.

Phase 3 — Citation Earning and Authority Signals (Weeks 5–10)

Restructuring earned citations on queries where the brand already had some relevance. The next layer — being cited for queries where competitors currently dominated — required authority signals that models use to decide who is credible.

External corroboration. Answer engines weight brands that are discussed in places they trust. We ran a focused digital-PR and partner-listing effort to place the company in relevant Pakistani business and HR communities: SME resource pages, Chamber of Commerce directories, HR-practitioner communities, and integration partner directories. The goal was not link equity in the classical sense; it was the density and diversity of independent mentions that a model can surface when answering “which tools do Pakistani SMEs use for payroll.”

A specific tactical choice mattered here: we prioritised corroborating sources that themselves get cited. A mention on a generic directory few models trust moves nothing; a mention in a well-maintained industry resource that already appears in AI answers cascades quickly. We tracked not just the count of placements but whether the placing pages themselves showed up as citations, and we weighted effort toward the latter. Twelve well-chosen corroborating mentions outperformed sixty low-quality ones, which is the opposite of the volume logic that governs old-school directory link building.

Source-of-truth technical content. For the 20 problem queries in the tracked set (“how to calculate EOBI contribution,” “how to file monthly tax withholding”), we built deep, authoritative explainers — the kind of page a model reaches for when it needs a citation to substantiate a claim. These pages linked clearly back to the product as the implementation path, but they earned their place by being the best available answer to the underlying question, not by selling.

Review and reputation surface. We consolidated public reviews onto a structured page (without inventing ratings — only aggregating genuinely collected customer feedback) and connected the entity so that when a model assembled an answer it could pull consistent, verifiable signals about customer base and traction.

Phase 3 results (by Week 10):

• Citation share moved from 19% to 36%
• Brand mentioned (cited or named) in 54% of tracked answers, up from 21%
• Three comparison pages became regularly cited sources in their own right

Phase 4 — Measurement and Compounding (Weeks 8–14)

The final phase turned the engagement from a project into a system the team could run themselves.

Attributing AI-sourced demand. We instrumented tracking for the demand AI visibility was actually generating — monitoring branded-search lift, “how did you hear about us” survey responses, and referral paths from AI surfaces — and built a shared dashboard so the head of growth could report AI-sourced pipeline to the board in the same breath as paid and organic.

Citation-monitoring cadence. The weekly query-set scoring from Phase 1 became a permanent operating rhythm. When a citation dropped, the team knew within a week and could restructure the relevant page before the loss compounded.

Compounding loop. As more pages were restructured and more external corroboration accumulated, new citations came faster and with less effort per query — the same dynamic that makes classic SEO compound, applied to answer engines.

Final Results at 90 Days

Metric	Before	After	Change
Priority queries cited in AI answers	8%	41%	+33 pts
Brand mentioned (cited or named) in answers	21%	54%	+33 pts
Non-branded organic clicks (classic search)	Baseline	+47%	Compounding
AI-search-sourced demo requests	Untracked	3.1x baseline	Instrumented
Branded search volume	Baseline	+38%	Recall effect
Commercial pages with extractable answers	9 of 60	57 of 60	+48 pages
Content-to-demo conversion (rewritten pages)	Baseline	+22%	Clarity effect

These are illustrative outcome ranges, built from the patterns we see across Pakistani B2B SaaS companies, not audited third-party figures. They exist to help a buyer sanity-check whether this kind of engagement fits their situation.

What Made This Work

1. Measurement preceded execution. The single most important move was building the citation baseline before touching any content. Until the company could see it was cited 8% of the time, AI visibility felt like a vague threat rather than a tractable problem with a scoreboard.

2. Answer-first beat answer-adjacent. The biggest citation gains came from restructuring existing pages, not from publishing net-new content. Pages that already ranked well classically but were written as essays jumped in citation share the moment they led with a clean, extractable answer.

3. Access was a silent tax. A meaningful share of the brand’s invisibility had nothing to do with content quality — AI crawlers were being blocked at the edge. Fixing crawl access was a one-week task with outsized returns, and it is the most common hidden defect in Pakistani SaaS stacks we audit.

4. Authority compounds. The first citations were hard and slow. Once the brand was established as a credible source on a topic cluster, subsequent citations in that cluster came faster — the same compounding dynamic every SEO recognises, now playing out inside answer engines.

5. Neutral comparison content out-pulled promotional content. Pages that honestly compared the product against competitors — including where competitors were stronger — were cited far more often than pages that only argued for the brand. Answer engines penalise self-serving pages and reward comprehensive ones.

6. Classic SEO and answer-engine optimisation reinforced each other. The team treated this as a fork — invest in GEO or invest in SEO. In practice the same answer-first structure, the same schema, and the same authoritative explainers moved both rankings and citations. Almost nothing in this engagement traded off classic search performance; the +47% non-branded organic lift came from the same edits that won citations.

What Teams Can Apply

For Pakistani SaaS and tech companies that suspect they are invisible in AI answers:

1. Measure first. Pick 30 to 60 priority commercial queries, run them weekly against the major AI surfaces, and score citation share. You cannot manage what you have not baselined, and the number is almost always lower than the team assumes.
2. Make your existing pages extractable. Before writing anything new, rebuild your top commercial pages so the lead answer to the buyer’s question sits in the first 60 words, supported by schema. This is the highest-return, lowest-effort move.
3. Open crawl access deliberately. Check server logs for AI crawler response codes. If you are returning 403 or 429 to the models generating the answers your buyers read, no amount of content work will compensate.
4. Earn citation through corroboration, not claims. Density and diversity of independent mentions — directories, partner pages, communities, neutral comparisons — matter more than self-description. Build the footprint a model can verify.
5. Instrument the demand. Track branded-search lift and “how did you hear about us” responses so AI-sourced pipeline becomes visible. What gets measured gets budgeted.

WeProms Digital has applied this framework across Pakistani vertical SaaS, fintech, developer-tools, and service businesses. The query sets, content architecture, and authority tactics change with each vertical — the measurement-first, answer-first, access-first sequence stays consistent.