Harvesting Question-Based Keywords from Quora’s Topic Feeds Using Graph Traversal

Most keyword researchers still treat seed expansion like a one-dimensional ladder – start with a head term, step to a few modifiers, then pray for long-tail orphans. That approach leaves an enormous reservoir of intent-signaling questions untouched, especially those buried deep inside community-driven platforms where real humans ask unfiltered queries. Quora’s topic feeds are a goldmine of question-based long-tail phrases, but scraping them linearly yields noisy, shallow results. The savvy play is to treat each topic as a node in an implicit graph, then traverse the feed’s relational structure using breadth-first search on user-generated subtopics and cross‑linked “Related Questions” to surface lexical variants that traditional keyword tools never index.

Why Quora specifically? Because the platform’s recommendation engine already clusters questions by semantic similarity, not just lexical matching. When you land on a topic feed – say “/topic/Organic-Chemistry” – you see the canonical question, but also a sidebar of “Related Topics” like “Organic Chemistry Reactions” or “Aromatic Compounds.” Each of those sub‑topics is itself a node that contains hundreds of questions, many of which are phrased as natural-language queries that Google’s BERT models reward. More importantly, Quora’s “Related Questions” widget underneath each answer isn’t random; it’s driven by co‑occurrence patterns among users who upvoted or viewed both questions. That co‑occurrence graph is a proxy for search intent clusters. A straightforward scraper that only grabs the top ten questions from a topic page misses the cross‑linked questions that live two or three hops away.

The execution requires a two-layer crawl. First, for a given seed topic, collect every “Related Topic” link from the topic page’s sidebar. These are often lowercase, hyphenated slugs. Second, for each of those sub‑topics, request the feed endpoint (typically `https://www.quora.com/topic//questions`). Parse the HTML to extract each question’s full text, but do not stop at the first page. Quora lazy-loads questions, so you need to simulate infinite scroll by sending GET requests with a `?offset=` parameter. Collect until you hit the “no more questions” signal or a practical limit of 500 questions per sub‑topic. This gives you a raw set of question strings. The real value appears when you deduplicate across sub‑topics using fuzzy string matching (Levenshtein distance of 0.8 as a threshold) and then run a syntactic parser to tag questions by pattern: “how to,” “what is,” “why does,” “best way to,” “difference between.” Those pattern tags become your seed clusters for content targeting.

But raw question text is still noisy. Many Quora questions contain fluff like “I’m a beginner, please help” or “This might be a stupid question.” Strip those by filtering against a curated list of introductory phrases. More important: enrich each question with a “search volume proxy” by cross-referencing the question’s text against Google’s “People Also Ask” responses for a sample of the head terms extracted from the question. You can do this programmatically via the PAA API endpoint (undocumented but stable) by building a head term from the first three content words of the question. For example, “How do I calculate the equilibrium constant for a substitution reaction?” yields head term “calculate equilibrium constant” – feed that into a PAA call and note which related questions overlap with your Quora harvest. The overlap ratio becomes a “intent alignment score.” Focus your content creation on questions that score above 0.3, meaning at least 30% of the associated PAA queries match your Quora-derived phrasing.

One pitfall: Quora’s topic feed is not static. Questions are added, removed, or merged daily, so your graph traversal should run on a weekly cron job, exporting a delta of new questions since your last crawl. Store the resulting question–subtopic graph in a PostgreSQL table with a custom `ltree` column to represent the hierarchical path from your seed topic to each leaf question. That lets you run queries like “find all questions that are two hops from ‘SEO Tools’ but not directly in ‘Keyword Research’” – perfect for uncovering adjacent intent clusters that competitors ignore.

Finally, don’t treat these questions as standalone keywords. Cluster them into “intent families” by applying topic modeling (LDA with 10–15 topics) on the aggregated question corpus from a single seed topic. Each family typically reveals a distinct stage in the user journey. For instance, under “Machine Learning,” one topic family might be “how to implement X algorithm” (practical implementation intent) while another is “why does model Y perform better than Z” (evaluation/decision intent). Build content that directly answers the question form, then interlink those pages using the graph relationships you discovered. That creates an internal linking structure that mirrors the actual query journey users take when clicking through Quora’s recommendation engine. The result is a keyword ecosystem that feels natural to both users and crawlers, because it was extracted from a platform that already optimizes for the very conversational queries Google now rewards.