AI Search Citations for a Baby-Care Ecommerce Store

The Client

A Lahore-based specialty retailer of baby and mother-care products, founded in 2020, selling about 430 SKUs across feeding (bottles, sterilizers, breast pumps), sleep (swaddles and sleep sacks), bath and skincare, cloth and eco disposables, carriers and strollers, and postnatal recovery for mothers. Revenue ran near PKR 14 million a month online at an average order value around PKR 4,800 — gear-heavy baskets — split between the brand’s own Shopify storefront, a Daraz flagship store, and a steady stream of Instagram DM orders handled by a three-person social team.

The customer base was exactly who GEO is for: first-time parents, anxious, researching at 2 a.m., asking precise questions. A post-purchase survey the brand ran over four hundred responses found that 22% had asked ChatGPT, Gemini, or Perplexity a product question before their first order — “which bottle for a colicky newborn”, “is talc-based powder safe”, “carrier for a Lahore summer”. When the founders ran those same questions themselves, the answers cited foreign parenting blogs with dollar prices, gray-market importers on marketplaces, and the occasional local listicle. Their store — with certified stock, local warranty, and PKR prices — never appeared.

Paid social carried the business, at 46% of revenue, and organic had been flat at around 15% share for three quarters. This case study is an illustrative composite: an anonymised engagement assembled from the patterns WeProms sees in Pakistani ecommerce GEO work, published so a growth team can judge the framework and the outcome shape against their own market.

The Problem

The diagnostic surfaced six blockers, and — unusually — crawl access was not one of them. The brand’s robots.txt already admitted GPTBot, ClaudeBot, PerplexityBot, and Google-Extended. The site was readable; it simply was not quotable.

• Near-zero citation presence. Across a fixed set of 80 questions run weekly against four AI surfaces, the brand was cited in about 6% of answers and mentioned without attribution in another 5%. For best-of and comparison questions, the citation rate was effectively zero.
• The citation layer was winnable but occupied. Of the sources engines did cite, 44% were foreign parenting blogs with no Pakistan pricing, 25% were marketplace listings (including gray-market importers), 19% were local listicles, and 12% were forums. Almost nothing cited was Pakistan-specific, structured, and current.
• Prose that engines cannot lift. The blog carried six lifestyle essays — gift guides, birth-story features — with no verdict paragraphs, no comparison tables, no FAQ pairs, nothing in the compact, self-contained shape an assistant reuses.
• Thin machine-readable signals. Only 31% of the 51 priority product and category pages carried valid Product, Offer, and FAQPage markup. Prices, availability, and age-range suitability were rendered visually, not structured.
• A fragmented entity. The brand operated under three spellings — a shortened trade name on Instagram, the full registered name on Daraz, a stylised version on Facebook — with no consistent Organization node or sameAs wiring, so engines could not consolidate them into one entity.
• No measurement. AI-referred sessions landed in direct and referral buckets. Leadership could not see the channel, so it could not be prioritised.

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

Everything was built on a question set, because GEO is won or lost on answering what buyers actually ask.

Assembling the 80 questions. We drew from four sources in priority order: the brand’s on-site search logs (eleven months), the Q&A sections of competitor marketplace listings, transcripts from the WhatsApp support line — by far the richest, because support agents answer real questions in real language all day — and Google Search Console queries. The set was tiered into three intents: safety and trust questions (“is X safe for newborns”), comparison questions (“glass vs plastic bottles”, “carrier vs stroller for a first-born”), and localised best-of questions (“best baby soap in Pakistan”, “sterilizer under PKR 15,000”). Roughly a third of the set was phrased in Roman Urdu or mixed Urdu-English, mirroring how the customer base actually types.

The weekly sweep. Each question ran against ChatGPT with search enabled, Perplexity, Gemini, and Google AI Overviews, with answers logged in a tracking sheet that separated citations (named, attributed, linked) from bare mentions. The baseline scoreboard: 6% citation, 11% mention. We also recorded which source won each question, which is what made the opportunity legible — the currently cited material was mostly non-Pakistani, unpriced in rupees, and often two years stale.

Baseline table:

Signal	Baseline
Citation rate, 80 questions × 4 surfaces	6%
Mention rate (unattributed)	11%
Questions where a Pakistan-specific source was cited	18%
Priority pages with valid schema	31%
Answer-format pages on site (verdict + table + FAQ)	0
AI-referred sessions (grouped)	~240/month

The audit set the thesis for the engagement: this was not an access problem or a domain-authority problem. It was an extractability problem — the brand had nothing on-site shaped like an answer.

Phase 2 — Answer Architecture and Schema (Weeks 3-5)

The content architecture. We structured the site into six topic hubs — feeding, sleep, bath and skincare, diapering, gear and travel, and mother recovery — and planned 34 answer pages beneath them: 18 comparison or best-of pages and 16 safety/how-to pages. Every page followed one discipline, applied in this order:

1. A 40-to-60-word verdict paragraph that answers the question outright — options named, PKR prices included, the deciding factor stated.
2. A comparison table carrying the attributes buyers actually weigh.
3. A short, self-contained rationale per option — sentences that stand alone if lifted.
4. An FAQ block of three to five follow-up questions.

The comparison table is where the local advantage compounds, because it carries what foreign sources cannot: rupee prices, parts availability in Pakistan, and voltage or climate suitability. An excerpt from the newborn-bottle page shows the shape:

Option	Price (PKR)	Material	Anti-colic	Spare parts in Pakistan	Best for
Wide-neck PES bottle	1,450	BPA-free plastic	Yes	Nipples widely stocked	Everyday use
Narrow glass bottle	2,200	Borosilicate glass	No	Order-in, 3-4 days	Purity-focused buyers
Anti-colic PP bottle	2,850	BPA-free plastic	Yes	Full set stocked	Colicky newborns
Wide-neck glass bottle	3,900	Borosilicate glass	Yes	Order-in, 3-4 days	Long-term, one-time spend

Safety pages were written to a stricter standard. For a category where buyers ask “is this safe for my baby”, marketing copy is a liability. Each safety page stated only what manufacturer specifications and public health guidance support, cited both, and closed with a clear line directing parents to their paediatrician for medical questions. The tone is a feature, not caution for its own sake: assistants selecting a source for safety-adjacent answers favour calm, corroborated, non-promotional material.

Roman Urdu FAQ blocks. For the twelve highest-traffic questions, a Roman Urdu version of the FAQ was published beneath the English block — questions like “glass bottle better hai ya plastic newborn ke liye?” — mirroring how a large share of the customer base phrases both searches and assistant prompts. These blocks began earning Urdu-phrased AI Overview citations within weeks, and they doubled as the raw material the support team pastes into WhatsApp replies.

Schema, validated not just deployed. Product, Offer, AggregateRating, and FAQPage markup went onto all 51 priority pages, BreadcrumbList across the hub structure, and priceCurrency set to PKR on every Offer node. The discipline was validation: each page had to pass the Rich Results test cleanly before it counted as done, taking priority-page coverage from 31% to 100%. This is the same schema groundwork described in our generative engine optimization engagement scope — markup is what lets an engine re-state your price and availability with confidence.

Phase 3 — Entity Consolidation and Distribution (Weeks 4-8)

One name, everywhere. The brand standardised on its full registered name across Daraz, Instagram, Facebook, and its Google Business Profile, kept the short form as a redirecting alias, and published a single Organization node on the storefront with stable logo, address, and sameAs links to every profile. A brief “checked by our product team” byline policy — every guide reviewed against manufacturer specs by a named member of staff — gave the engines a person to attach to the publisher entity without inventing clinical credentials the brand does not have.

Corroboration, not link farming. Distribution stayed modest and deliberate: two guest buying guides placed in Pakistani parenting publications, one product-safety roundup contribution, and a post-delivery WhatsApp review flow that added 340+ verified buyer reviews to priority product pages, feeding genuine AggregateRating signals. We avoided directory links entirely; coherence and corroboration appear to matter more to citation selection than raw link counts.

Wiring the catalog into the answers. Every answer page linked down to the specific SKUs it discussed with descriptive anchors, every product page linked up to its hub, and an llms.txt index pointed crawling engines at the hub layer. By week 8 all 34 answer pages were indexed, and the six hubs had become the entry points engines hit first.

The answer layer also did double duty in classic search. The FAQPage blocks earned rich-result eligibility on all 34 pages, and the comparison tables — the same assets built for AI extractability — turned out to be the best-converting merchandising the brand had published: click-through from the answer pages to the specific products they discussed ran near 18%, against roughly 9% on the older prose guides. Google’s AI Overviews drew from the same FAQ blocks, which meant one content investment served three surfaces — classic listings, AI Overviews, and standalone assistants — rather than three competing formats. That is the quiet economic argument for GEO in a market like Pakistan, where content budgets are thin: the answer-first format is not a separate channel to fund, it is a better version of the buying guide most ecommerce brands need to write anyway.

Phase 4 — Measure, Refresh, and Compound (Weeks 8-12)

The weekly sweep became the operating instrument, the discipline behind our AI search visibility monitoring engagements. Pages that earned citations were refreshed monthly — prices first, because PKR volatility means a stale price is a wrong price, and a page engines learn is stale stops being cited. Pages with no citation after six weeks were restructured; in nearly every case the missing element was the comparison table or a verdict paragraph that had drifted past 60 words.

We also grew the tracked set from 80 to 130 questions at week 10 to guard against overfitting — the citation rate held on unseen questions at 31%, which is the more honest number and the reason we report a range rather than a single victory. In GA4, chatgpt.com, perplexity.ai, and gemini referrals were grouped into a single “AI assistants” channel (Google AI Overview referrals are tagged where GA4 exposes them, with the caveat that they remain partly invisible inside google.com sessions). The channel converted at 3.1% against 2.7% for organic overall — small samples early, but directionally consistent with buyers arriving pre-qualified by an answer that already named their option and its price.

Final Results

Measured at 90 days against the Phase 1 baseline:

Metric	Before	After	Change
AI answer citation rate (80 questions, 4 surfaces)	6%	36%	+30 pts
Mention rate (unattributed)	11%	49%	+38 pts
Answer pages with at least one citation	3 of 34	22 of 34	+19
Non-branded organic sessions	Baseline	+34%	Compounding
AI-referred sessions per month	~240	~2,600	~11x
Branded search impressions (QoQ)	Baseline	+27%	Citation halo
Organic revenue share	15%	24%	+9 pts
Priority pages with valid schema	31%	100%	+69 pts

Each line traces to a phase: the citation and mention lift to the answer architecture and comparison tables of Phase 2, sustained by the refresh loop in Phase 4; the branded-search halo to thousands of parents reading the brand’s name inside answers; organic sessions and revenue share to the schema, entity, and internal-linking groundwork; and the AI-referred session curve to the channel grouping that finally made the demand visible. These figures are illustrative — a realistic outcome shape for a mid-catalog Pakistani ecommerce brand in a question-heavy category, not an audited third-party result.

What Made This Work

1. Verdict-first pages with PKR tables. The 40-to-60-word verdict and the comparison table earned the overwhelming majority of citations. Engines reconstruct answers from compact, structured sources; rupee prices and parts-availability columns are attributes foreign blogs structurally cannot supply.
2. The question set came from our own channels. On-site search, marketplace Q&A, and WhatsApp support transcripts produced questions in the exact language buyers use — including Roman Urdu. A translated foreign keyword list would have aimed the content at queries nobody here types.
3. Schema and entity work made citations land on the brand. Valid markup let engines re-state price and availability accurately, and the single consolidated entity meant the credit went to the brand’s domain, not to a marketplace listing or a similarly named seller.
4. The weekly sweep decided where effort went. Per-engine, per-page citation tracking turned GEO from a content-volume game into a loop: cite → refresh → corroborate, or restructure. When the set expanded to 130 questions, the 31% hold rate confirmed the result was a capability, not a memorised test.
5. The safety tone matched the category. Calm pages built on manufacturer specs and public-health guidance, with a clear line to consult a paediatrician, were citable on trust-sensitive questions in a way promotional copy never is.

What Teams Can Apply

For Pakistani ecommerce brands in question-heavy categories — baby care, health, electronics, beauty:

1. Mine your own channels for the question set before writing a word. Your on-site search logs, marketplace Q&A, and support transcripts contain the exact phrasing of your buyers’ questions, Urdu and English both. That language, not a translated keyword list, is what you should be publishing answers in.
2. Write the 40-to-60-word verdict before anything else on the page. If you cannot answer the question in three sentences with a price attached, the page is not ready. The table, rationale, and FAQ all reinforce that first block — they do not replace it.
3. Put PKR prices and local availability in the table. It is the one structural advantage a Pakistani retailer has over every foreign source an assistant currently cites. Prices move with the rupee, so refresh monthly or lose the citation.
4. Consolidate the entity before chasing links. One name spelling everywhere, one Organization node, sameAs wiring, and validated Product schema are what turn a citation into branded traffic instead of a generic mention. A handful of genuine corroborating placements beats any volume of directory links.
5. Group AI referrals in analytics now, before you need the data. A named channel with a baseline takes an hour to set up and is the difference between arguing about AI search and budgeting for it.

The buyer-side shift this engagement was built for is covered in our digital marketing for ecommerce hub, and the tracking side — question sets, weekly sweeps, citation-versus-mention separation — is the operating habit worth importing first.