AI Answer Citations for an Ecommerce Brand

The Client

A direct-to-consumer home and kitchen brand based in Multan, selling cookware, bakeware, kitchen organizers, and small appliances through their own storefront alongside a marketplace listing. The catalog sat at roughly 650 SKUs across 18 collections, monthly revenue hovered around PKR 18 million, and repeat-purchase behavior across Punjab was genuinely strong — the kind of brand that wins once a customer has held the product.

The problem was everything upstream of that first purchase. The founder had started asking ChatGPT, Perplexity, Gemini, and Google’s AI Overviews the same questions their customers were asking: “best non-stick cookware set in Pakistan”, “kitchen organizer for a small apartment in Lahore”, “induction-friendly pressure cooker under PKR 8,000”. Their brand never appeared. A couple of affiliate listicles and one better-funded competitor showed up almost every time.

This was no longer a fringe channel. The brand’s own post-purchase survey, run over a quiet two-week window, showed that roughly one in four new buyers had consulted an AI assistant before deciding. They were losing demand they could not see, to competitors whose product was not necessarily better but whose content was easier for an AI engine to quote.

The Problem

Five concrete blockers showed up in the diagnostic:

1. Near-zero AI visibility. Across a fixed set of 50 evaluative and category prompts run weekly against five engines, the brand was cited in roughly 4% of answers. For “best for” and comparison prompts the rate was effectively zero.
2. Blocked AI crawlers. A previous developer had copied a restrictive robots.txt that disallowed GPTBot, ClaudeBot, PerplexityBot, anthropic-ai, and Google-Extended. The catalog was invisible to the crawlers that feed generative answers, while aggressive scrapers were still getting through.
3. Thin machine-readable signals. Only 22% of priority product and category pages carried valid Product, Offer, and FAQPage markup. Prices, availability, and ratings were rendered visually but not structured for extraction.
4. No quotable answer content. The site had product pages and a short blog. There were no comparison tables, no “best X for Y” pages, and no paragraphs that directly answered a question in the compact, self-contained shape an AI engine will lift verbatim.
5. Weak entity footprint. The brand was not recognized as a distinct entity. There was no consistent Organization and sameAs wiring, no Knowledge Panel, and no reference footprint the engines could use to disambiguate the brand from similarly named sellers.

Non-branded organic traffic had been flat for three quarters and organic revenue sat around 14% of total. Paid acquisition was carrying the weight, and the team suspected, correctly, that the buyers they were paying to reach on Meta and Google were the same buyers who would have found them through an AI answer for free.

Phase 1 — Answer-Landscape Audit and AI Access Cleanup (Weeks 1-2)

The first two weeks were about understanding the battlefield and unlocking the door.

Building the query set. We assembled 50 prompts that reflected real Pakistani buying language across three tiers: broad category prompts (“best cookware sets in Pakistan”), specific evaluative prompts (“induction-compatible non-stick pan that is oven-safe”), and localized prompts (“kitchen organizer delivery to Multan”). The set was built from three sources — the brand’s own search-console and on-site search data, a scrape of competitor FAQ sections, and a panel of prompts we ran in Roman Urdu and mixed Urdu-English, because that is how a meaningful share of buyers actually query assistants. Each prompt was run weekly against ChatGPT, Perplexity, Gemini, Google AI Overviews, and Copilot, with the answers logged in a tracking sheet.

A distinction mattered here that most teams miss. A citation is a named link or attributed mention of the brand in an answer. A mention is the brand or product appearing without a link or clear attribution. The baseline scoreboard separated the two: a 4% citation rate and a 7% mention rate. Mentions are worth something — they confirm the brand is in the engine’s consideration set — but citations drive the referred traffic and the revenue attribution, so the program optimized specifically for citation rate, not raw visibility.

Crawl-access audit. The robots.txt review was the single highest-leverage fix in the engagement. The file was disallowing the exact crawlers that feed generative answers while doing little to stop actual scrape abuse. We rebuilt it to allow content-benefit AI crawlers (GPTBot, ClaudeBot, PerplexityBot, anthropic-ai, Google-Extended, CCBot under a polite crawl-delay) and block known aggressive scrapers by user-agent and path. We added an llms.txt index pointing engines at the category hubs and the buying-guide layer.

Index and canonical hygiene. A crawl of the storefront surfaced 140 canonical conflicts on collection variants and 60 category pages tagged noindex by a theme update. We resolved the canonicals and removed the stray noindex flags, then submitted a clean sitemap. This is unglamorous work, but an AI engine will not cite a page it cannot fetch or reconcile to a single canonical URL.

Baseline snapshot at end of Phase 1:

Signal	Before	After (Week 2)
AI crawlers allowed	0 of 5	5 of 5
Canonical conflicts	140	0
Stray noindex on category pages	60	0
Priority pages fetchable by bots	61%	100%

Phase 2 — Entity and Structured-Data Foundation (Weeks 3-5)

With the site fetchable, we made the catalog machine-readable.

Schema across the catalog. We implemented Product, Offer, AggregateRating, Review, FAQPage, and BreadcrumbList markup on every priority product and category page. Validation was the discipline here, not implementation: every priority page had to pass the Rich Results test cleanly. We hardened product attributes — material, dimensions, color, availability, priceCurrency set to PKR — because AI engines extract and re-state these fields directly. A product page that says “oven-safe to 220°C, induction-compatible base” in a structured field is far more citable than one that buries the same fact in a paragraph.

Entity building. We wired consistent organization signals: a single Organization node with a stable name, logo, NAP, and sameAs links to the marketplace store and social profiles, plus a founder profile linked from the about page. The goal was to give the engines enough rope to recognize the brand as a distinct entity rather than a keyword match.

Self-describing product pages. Each priority product page was rebuilt to stand alone as an answer source: a one-line definition of what the product is and who it is for, a compact spec block, three to five FAQ entries answering the exact questions buyers ask, and a related-products cluster with descriptive anchors. The spec block mattered more than expected — engines extract and re-state structured attributes far more reliably than prose. Before the rebuild, fewer than a third of priority products exposed key attributes in a machine-readable form; after, every priority page did.

Attribute coverage (priority pages)	Before	After
Material	38%	100%
Dimensions / capacity	44%	100%
Compatibility (induction / oven)	19%	100%
Price in PKR (structured)	61%	100%
Availability signal	52%	100%

This work is described further in our generative engine optimization engagement scope.

Internal answer clusters. Category hubs were restructured to carry an FAQ block and short answer paragraphs targeting the questions the audit had surfaced, with product pages linking back up to the hub. This gave engines a clean path from a broad question down to specific products.

Phase 3 — Answer-Content Build at Scale (Weeks 4-9)

Schema makes a page readable; answer content makes it quotable. This phase produced the assets the engines actually cite.

Comparison and “best for” pages. We shipped 24 pages built around how buyers ask AI assistants for recommendations: “best cookware for induction cooktops in Pakistan”, “kitchen organizers for small apartments”, “pressure cookers under PKR 8,000”. Each page followed a deliberate structure:

• A 40-60 word lead paragraph that answers the question directly, naming the top options and the deciding factor.
• A comparison table with the attributes buyers actually weigh — price in PKR, material, induction compatibility, warranty, capacity.
• A short rationale per option, written in self-contained sentences an engine can lift.
• An FAQ block answering the follow-up questions the audit showed up in prompts.

Comparison tables turned out to be the highest-citation format by a wide margin. Engines prefer structured, scannable content they can reconstruct with attribution, and a clean table with PKR prices and clear attributes is close to ideal source material. An illustrative excerpt from one of the winning pages shows the shape that earned citations:

Option	Price (PKR)	Material	Induction	Capacity	Best for
Model A	6,400	Aluminium non-stick	Yes	24 cm	Everyday cooking
Model B	9,200	Stainless tri-ply	Yes	22 cm	Heavy use, longevity
Model C	4,800	Ceramic non-stick	No	26 cm	Budget, low-oil cooking

When an engine is asked “best induction cookware under PKR 10,000 in Pakistan”, a table like this is the cleanest possible source — it answers the question, names options with PKR prices, and the attributes map directly to the filters in the prompt. Pages built in this shape earned citations at roughly three times the rate of prose-only equivalents in the same topic cluster.

Product description refresh. Sixty priority product descriptions were rewritten with quotable specification sentences — “ceramic non-stick coating, oven-safe to 220°C, induction-compatible base, 24 cm diameter” — replacing marketing prose that an engine could not reliably extract. We left brand voice intact in the hero copy; we just made the technical facts structured and self-contained.

FAQPage coverage. Every category hub and every comparison page carried a FAQPage block answering the three to five questions the query audit had surfaced for that topic. These blocks did double duty: they earned rich-result eligibility in classic search and they gave AI engines ready-made question-and-answer pairs to synthesize.

Phase 4 — Citation Monitoring and Compounding (Weeks 8-16)

The final phase turned a one-time build into a compounding channel by closing the measurement loop.

Citation tracking. The 50-prompt weekly sweep was expanded to 120 prompts and instrumented as a proper monitoring layer, tracking citation rate, mention rate, and sentiment per engine. This is the same discipline behind our AI search visibility monitoring work. Knowing which pages earned citations — and which did not — told us where to double down.

Attribution of AI-referred traffic. We grouped AI platforms (ChatGPT, Perplexity, Gemini, Copilot, and AI Overview referrals) into a single channel in GA4 and attributed revenue to it. AI-referred sessions scaled from roughly 120 a month to 3,400 a month over the engagement, with a conversion rate in line with organic search once the traffic was large enough to measure reliably.

Refresh and prune. Pages that earned citations were refreshed quarterly with current prices, new variants, and updated FAQ answers. Pages that did not earn citations after eight weeks were either restructured (usually missing the comparison table) or consolidated. The comparison-table format was systematized across every new page.

Supporting authority signals. A small number of targeted digital PR placements and an aggressive review-generation program strengthened the entity and gave the engines more corroborating signals. We avoided low-quality directory links entirely — AI engines appear to weight coherence and corroboration more than raw link count, and a messy backlink profile can actively hurt citation likelihood.

Final Results at Five Months

Metric	Before	After	Change
AI answer citation rate (50 queries, 5 engines)	4%	36%	+32 pts
Non-branded organic sessions	Baseline	+58%	Compounding
AI-referred sessions per month	~120	~3,400	~28x
Organic revenue share	14%	26%	+12 pts
Priority pages with valid schema	22%	100%	+78 pts
Referring domains	41	69	+28
Indexed comparison and best-for pages	6	30	+24

Every number in the table traces back to a specific phase: the citation-rate jump to the comparison pages and crawl-access fix in Phases 1 and 3, the organic lift to the entity and schema foundation in Phase 2, and the AI-referred traffic to the monitoring and refresh loop in Phase 4. The shape of the outcome — citations leading, organic compounding, paid dependency falling — is the realistic pattern for a mid-catalog D2C brand investing in GEO over a single quarter.

What Made This Work

Four factors drove the result, in roughly this order of leverage:

1. Unlocking crawl access first. The brand was literally invisible to the crawlers that feed generative answers. No content investment would have earned citations while GPTBot and Google-Extended were disallowed. The robots.txt fix was a five-minute change with outsized impact, and it is the first thing to check on any GEO engagement.
2. Quotable structure beat quotable prose. Comparison tables and 40-60 word answer leads earned the majority of citations. AI engines reconstruct answers from structured, scannable source material with clear attributes. A beautifully written paragraph that buries the deciding factor will lose to an ugly table with PKR prices every time.
3. Schema and entity work made content extractable. Markup gave engines reliable price, availability, and attribute fields to re-state, and the entity wiring let them attribute the citation to the right brand. Without it, even cited content was often attributed to the marketplace seller or to a generic category.
4. The measurement loop focused the spend. Because citations were tracked weekly per engine and per page, we knew exactly which pages to refresh and which to consolidate. GEO without measurement is guesswork; the 120-prompt sweep was the single most important operational asset by the end.

What Teams Can Apply

For a Pakistani ecommerce brand that suspects it is invisible in AI answers:

1. Audit robots.txt for AI crawlers before anything else. If GPTBot, Google-Extended, and PerplexityBot are disallowed, your catalog is invisible to generative answers regardless of content quality. This is a minutes-to-fix issue that blocks everything else.
2. Write answer-first content with comparison tables. For every category, publish at least one “best X for Y” page that leads with a direct answer and includes a comparison table with PKR prices and the attributes buyers actually weigh. This is the format engines cite most.
3. Add Product, Offer, and FAQ schema, and validate it. Structured price, availability, and attributes are what engines re-state in answers. Run the Rich Results test on every priority page; half-built markup does not help.
4. Track AI citations as a channel. Run a fixed weekly prompt set against the engines that matter to your buyers, group AI-referred sessions in analytics, and attribute revenue. You cannot optimize what you do not measure, and the cost of the tracking is trivial next to the spend it informs.

This illustrative engagement reflects the patterns WeProms Digital sees across Pakistani D2C brands investing in generative engine visibility — the specific query sets, comparison angles, and schema priorities change with each vertical, but the crawl-access, structured-content, and measurement sequence stays consistent. The AI search citations work for Pakistani ecommerce covers the broader buyer-side trend this engagement was built to address.