Generative Engine Optimization for a Pakistan SaaS

The Client

The client is a Sialkot-based SaaS company selling HR and payroll software to Pakistani manufacturers — payroll runs, attendance, leave records, EOBI registration and monthly filings, and gratuity calculations for companies with 80 to 500 employees. It was founded in 2019 by two brothers from a garment-export family who had watched their own suppliers process factory payroll in spreadsheets and WhatsApp groups, and it had grown to roughly 190 paying companies, concentrated in Punjab’s manufacturing belts around Sialkot, Faisalabad, and Lahore, at about PKR 340 per employee per month.

The company is exactly the kind of business that grows on trust and never quite gets around to demand generation. A 24-person team, two of them in any kind of marketing role, with acquisition running on referrals, the founders’ LinkedIn presence, and occasional word-of-mouth from accountants. Classic SEO had been attempted twice and abandoned both times — a blog of generic HR posts that drew nothing, and an agency retainer that produced traffic without demos. Sialkot itself shaped the go-to-market: the cost base and the founders’ manufacturing network made Punjab’s factory belt the natural territory, which meant the queries that mattered were Pakistan-specific by definition — a fact that later worked in the engagement’s favour, because Pakistan-specific prompts are exactly where generic international content does not satisfy the question.

The trigger for this engagement came from a lost deal, in the way these things usually do. In mid-2026 a Faisalabad textile mill the founders had courted for four months chose an international suite instead. The finance director’s parting feedback was relayed almost verbatim: “I asked ChatGPT for the best payroll software for companies in Pakistan, and you weren’t on the list.” The founders tried the same prompt that evening and found their category dominated by global brands they never compete against in an actual sales process, plus one Karachi competitor — and their own product absent. They came to us with a specific question: how do we get into those answers?

The Problem

The baseline audit answered a blunter question first: the brand was not merely ranking poorly in AI answers — for most of the previous year, the leading AI systems could not read the site at all.

• Citation coverage was 3 of 60 queries (5%). We built a priority set of 60 commercial queries across four engines — ChatGPT, Gemini, Perplexity, and Google AI Overviews — covering category searches (“best payroll software Pakistan”), feature searches (“payroll software with EOBI integration”), comparison searches (“payroll software vs accountant”), and vertical searches (“payroll software for textile mills”). The brand appeared in three, one of them its own near-branded name.
• A security plugin had been blocking AI crawlers since 2025. A blanket “block bad bots” firewall rule, installed during a spam wave, matched GPTBot, ClaudeBot, and PerplexityBot-class agents as scrapers and returned challenge pages. Google’s crawler was allowed; nearly every AI engine’s was not. Nobody had ever connected the rule to search visibility, because nothing broke audibly.
• Entity data contradicted itself. A 2021 rebrand left the old company name live on a dozen directories, two profiles listed the firm under “software house” and another under “HR consultancy,” and the founders’ bylines pointed at both names. Models assembling an answer from public data found no consistent story to verify.
• The site offered nothing citable. Pricing sat behind “contact sales.” Feature pages said the product was “reliable, scalable, and trusted” — language a model cannot verify or quote. There was no data, no specifics, and structured data on only four pages.
• Competitors filled the vacuum. International suites dominated the generic answers, and the one Karachi competitor — with public pricing and active G2 reviews — appeared in the Pakistan-specific ones. SaaS companies in Pakistan selling high-consideration products face this pattern acutely: the buyer asks an AI assistant for a shortlist, the assistant names whoever it can verify, and the losing vendor never learns it was in the room.

The strategic framing we agreed on with the founders: AI answers are not a channel you buy your way into, and they are not classic SEO either. They reward being readable, quotable, and corroborated. That became the four-phase plan.

Phase 1 — Citation-Gap Audit and Entity Baseline (Weeks 1-2)

The first fortnight produced two artifacts that carried the whole engagement.

The first was the query grid: 60 commercial prompts, winnowed from 140 candidates by scoring actual sales-call language — the questions prospects ask on discovery calls, mapped to how a buyer would phrase them to an assistant. The grid ran weekly across the four engines, logged in a shared tracker, with the brand’s presence recorded as cited / mentioned / absent, plus which URL the engines quoted when they did cite. This turned “AI visibility” from a vibe into a trend line, and it is the single practice we most insist on: without a fixed grid, every claim about AI visibility is an anecdote.

The second was the entity baseline: a census of everywhere the brand existed publicly — website, directories, G2, Capterra, LinkedIn company and founder profiles, Crunchbase-class listings, the old name’s remnants — with every inconsistency catalogued. The census found 17 properties needing repair and, buried in a server config review, the firewall rule that would otherwise have made every downstream fix pointless.

The audit also segmented the 57 uncited queries by why the brand was absent, which set the order of operations:

Gap type	Queries affected	Fix owner
Crawler blocked outright	All (site-wide)	Dev / firewall
Nothing citable on-page	31	Content + product
No comparison surface	14	Content
Entity unverifiable	9	Off-site repair

The lesson of the phase: sequence matters. Opening crawler access before publishing citable content would have made the site readable but thin; publishing content while crawlers stayed blocked would have been invisible labor. Both tracks opened together.

Phase 2 — Crawler Access, Schema, and Entity Repair (Weeks 3-5)

Phase 2 ran as a structured generative engine optimization build, and its most impactful change was one line of configuration: the firewall rule was narrowed to actual abuse traffic, robots.txt was rewritten to explicitly allow GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot, Google-Extended, and CCBot, and the site added an llms.txt index describing its key pages in plain text. The effect is unglamorous and decisive — for the first time in over a year, the engines building AI answers could read the product.

On-site structure followed. Organization schema was deployed with a complete sameAs graph linking the company’s verified profiles; SoftwareApplication schema carried the product’s category, features, and pricing; FAQ and HowTo markup went onto the pages rebuilt in Phase 3. Pages with valid structured data rose from 4 to 38 across the roughly 90-page site.

Then the entity repair, which is clerical work with outsized leverage. The 17 properties were reconciled to one name, one category, one founder story, one URL — directories corrected, stale profiles claimed or closed, the old domain properly redirecting, G2 and Capterra profiles completed with consistent positioning. Two weeks of tedious consistency work is what allows an engine to verify, from independent sources, that this company is what it says it is. Recommendation is downstream of verification.

The last act of the phase was commercial honesty: the pricing page was rebuilt around a public table — PKR per employee per month by tier, implementation fees, timeline — over the founders’ initial objection that competitors would see it. The counterargument, which held, was that buyers’ assistants were already comparing on price from competitors’ public pages, and the only seat at that table is a public one.

Phase 3 — Answer-Shaped Content and Corroboration (Weeks 4-8)

With the site readable and verifiable, Phase 3 made it quotable.

Fourteen commercial pages — the homepage, product, features, and key solution pages — were rewritten answer-first: question-form headings, a direct answer of roughly 40–60 words immediately under each, then depth. AI engines extract quotable statements the way a busy executive reads: headline, first sentences, tables. Long narrative intros, however well written, do not become citations.

Three new content surfaces were built for the grid’s biggest gap segments:

1. A comparison hub — honest, specific pages for “payroll software vs spreadsheets,” “vs outsourcing to an accountant,” and “vs international suites,” with feature tables that conceded where competitors genuinely won (multi-country payroll, for instance) and won where the product did (EOBI filing workflows, Urdu-language support, local bank integration). Models assembling “alternatives” answers reach for exactly this material.
2. Compliance references — six reference pages covering EOBI registration and contribution rules, gratuity calculation under Pakistani practice, payroll tax slabs, minimum-wage compliance, and year-end processing, written to be the clearest statement of the facts on the Pakistani web rather than a pitch.
3. Citable data — a small stats page publishing aggregates from their own customer base with permission: median payroll processing time before and after adoption, error rates in spreadsheet-era vs automated filings. This became the single most-quoted page within six weeks, because a model answering “how long does payroll take for a Pakistani manufacturer” prefers a number with a source over adjectives.

Corroboration then moved off-site. The founders’ review-drive lifted G2 reviews from 11 to 29 over the window, a founder contributed a bylined article to a Pakistani tech publication, and an exporter-association newsletter carried a second piece — two independent referring domains that also served as entity verification. None of this is link-building for PageRank in the classic sense; it is building the web of consistent, independent statements about who the company is that answer engines weigh before recommending anyone.

Phase 4 — Track, Iterate, and Compound (Weeks 8-12)

The final phase was measurement discipline under our AI search citation tracking process: the 60-query grid run weekly across all four engines, each movement logged with the citing URL and the phrase quoted.

The grid drove iteration rather than celebration. Where AI Overviews cited the brand but ChatGPT did not, we strengthened the FAQ schema and the directness of the opening answers — the engines read differently, and the grid showed it. Where a competitor held a query, a comparison page got sharper and more specific rather than broader. Where citations arrived but quoted the homepage instead of the pricing or comparison page, internal linking was adjusted to make the intended page the clearest answer. Eleven of the 25 cited queries at day 90 were won in this iteration loop after the initial build had plateaued at 14.

Attribution was instrumented honestly: GA4 segments for referrals from the AI platforms’ domains, UTM-tagged links in llms.txt, and a “how did you hear about us” field on the demo form that we treated as directional signal, not gospel. The point was to make AI-influenced pipeline visible at all, since the default state of AI referrals is to look like direct traffic.

The phase closed with enablement rather than a handover document. The two marketing staff were trained to run the weekly grid themselves — thirty minutes, four engines, one sheet — and to read the movement patterns the way Phase 4 had: engine gaps pointing at schema and directness, competitor holds pointing at comparison content, homepage-instead-of-page citations pointing at internal linking. The founders kept one rule for themselves: no new page ships without a citable fact on it. That single habit protects more of this engagement’s gains than any tool does.

Final Results

At day 90, against the week-2 baseline:

Metric	Before	Day 90	Change
Priority queries with ≥1 AI citation	3 of 60 (5%)	25 of 60 (41%)	+36 pts
Google AI Overviews citations	2	21	+19
ChatGPT citations	1	17	+16
Perplexity citations	0	14	+14
Gemini citations	0	9	+9
Organic sessions	9,400/mo	12,600/mo	+34%
AI-platform referral sessions	~180/mo	1,240/mo	~7x
Monthly demo requests	86	112	+30%
Branded search impressions	4,100/mo	7,300/mo	+78%

The citation math traces directly to the phases: crawler access and schema lifted coverage to 14 queries by week 7, and the Phase 4 iteration loop carried it to 25. Organic sessions grew 34% on the combination of AI Overviews placements, the comparison and reference pages ranking in classic search, and a branded-search lift as AI answers repeated the name. Demo requests rose 30%, with AI-referred visitors converting at close to the site average — the hypothesis being that a buyer who arrives from an AI answer has already been partially qualified by the citation itself. By the founders’ own pipeline attribution, roughly six new customers closed in the window with AI in the discovery path, worth about PKR 0.5 million in monthly recurring revenue — directional, not audited, and included because the question a founder actually asks is “did any of this produce revenue.”

What Made This Work

1. The crawler block was found first. A one-line firewall fix, in place for a year, had made every other optimization irrelevant. Checking whether AI engines can read your site takes ten minutes and belongs at the top of every AI-visibility audit.
2. Citable facts beat persuasive copy. Public PKR pricing, EOBI specifics, and processing-time data are quotable statements a model can verify and repeat. “Reliable, scalable, trusted” is unverifiable and therefore invisible to an answer engine.
3. Entity consistency did the verification work. Seventeen reconciled properties gave the engines independent corroboration of who the company is — the precondition for being recommended alongside global brands.
4. The fixed query grid made it a science. Sixty prompts, four engines, weekly cadence turned anecdotes into a trend line, showed which engine lagged, and drove eleven late citation wins that a launch-and-leave approach would have missed.
5. Comparison pages conceded honestly. Pages that admitted where international suites win read as credible to both buyers and models — and captured the “vs” and “alternatives” prompts where the real shortlisting happens.
6. Attribution was instrumented, not assumed. Referral segments, UTM-tagged llms.txt links, and the demo-form field made AI-influenced pipeline visible instead of vanishing into direct traffic.

What Teams Can Apply

1. Check crawler access before anything else. Look at robots.txt and any firewall or security-plugin bot rules for GPTBot, ClaudeBot, PerplexityBot, and Google-Extended. In Pakistan, aggressive anti-bot configurations are common — and they silently exclude you from the answers your buyers read.
2. Publish one citable facts page before you write ten blog posts. Pricing, compliance specifics, timelines, and first-party data are what engines quote. Most Pakistani B2B sites have none of it public.
3. Make your entity data identical everywhere. One name, one category, one story across your site, directories, review platforms, and founder profiles. Consistency is what lets a model risk recommending you.
4. Track a fixed weekly query grid. Define your 40–60 commercial prompts, run them across ChatGPT, Gemini, Perplexity, and AI Overviews weekly, and log movement. What gets measured gets cited.
5. Answer the comparison prompts honestly. Your buyers are already asking “X vs Y for Pakistan” — if you will not publish the honest comparison, the engines will assemble one from your competitors’ pages instead.