Loading page
Target outcome
An early-warning diagnosis with sources, severity, and remediation steps
Detect the anomaly
Trace the narrative
Assess severity
Contain the risk
Run in Claude and Run in ChatGPT open a new browser tab with the prompt already typed. Nothing is sent until you submit it.
Social listening waits for someone to publish. An assistant can change how it describes you with nothing published, and the shift lands in the answers you already monitor. This workflow puts a number on it and assigns a severity tier that decides the response.
Early means one thing here: earlier than a media alert, because nothing has to be published for the wording to move. It is not earlier than your monitoring cadence, and it sees nothing outside your monitored prompts, providers and markets.
The second gap is scale. A team stares at a curve, feels a drop, argues whether it counts. Nothing triggers without a threshold, and that threshold has to survive thin days: the daily rows carry the answer count, not how many answers scored a sentiment, so a quiet day swings the average alone.
1. Build the baseline, then the tiers. get_sentiment_trend returns one row per day with data: average sentiment out of 100, average position, answer count, mention rate. Baseline is the median daily sentiment, ignoring days below half your median volume. Tier 1 is one qualifying day 10 or more points under, tier 2 three consecutive such days or one day 20 points under, tier 3 a 20-point gap held across three.
2. Rule out the category. get_competitor_comparison gives your mentions, average sentiment, position and market rank against direct competitors; get_competitor_evolution gives daily figures, one competitor per call. If the leaders dip on the same dates, the movement belongs to the category and the tier drops by one.
3. Read the wording. get_answers lists answers by date with prompt text, provider, whether you were mentioned and whether your domain was cited, up to 100 rows per page, no sentiment field. get_answer_details adds the full text, competitor mentions with position and sentiment, cited URLs.
4. Trace the cited domains. get_source_profile takes a domain name or an id and returns its mention stats, most cited URLs and the competitors co-cited in the same answers. A framing problem becomes documents someone can read and sometimes correct.
5. Check the snapshot last. get_brand_perception returns the latest scores out of 100 with a SWOT list, and get_perception_history up to twenty monthly snapshots, with no date range. A threat naming what the answers say raises the tier by one.
Act on the tier, not the tone. Tier 1: record dates and baseline. Tier 2: open a ticket only when the same claim appears in at least two answers from at least two providers, because one provider repeating itself is one provider. Tier 3: escalate that day.
Two tier 2 findings in the same week: take the one seen on more providers, then the one whose cited source has someone you can write to.
Absence is a state, not zero. A null sentiment means no scored mention, missing self stats mean you were absent from the window, an empty answer sample means unresolved rather than clean.
Run an AI reputation check and give me a severity tier before advice.
Call list_brands. If more than one brand comes back, ask which to use before pulling data.
1. get_sentiment_trend over 90 days. Baseline is the median daily sentiment, ignoring days whose answer count is under half the median. A null sentiment means no scored mention, never zero. Tier 1: one qualifying day 10 or more points below baseline. Tier 2: three consecutive such days, or one day 20 points below. Tier 3: 20 points below across three qualifying days. Under 10 days with data, cap at tier 1 and say so.
2. get_competitor_comparison for the same window, then get_competitor_evolution for the top three direct competitors, one call each. If they fall on the same dates, lower the tier by one: the move is category wide. If no stats come back for my brand, report that I was absent from the window and treat it as tier 2.
3. get_answers on the affected dates, then get_answer_details on up to eight that mention me. Quote the wording a user would read, competitor positions and cited URLs. If nothing comes back, widen by a week and call the sample unresolved, not clean.
4. get_source_profile on the three domains cited most in that sample: mention counts, top URLs, co-cited competitors. Associated sources, not causes. Name any domain returning nothing as skipped.
5. get_brand_perception and get_perception_history. No snapshot means keep the tier from step 1. A weakness or threat naming the same issue as step 3 raises the tier by one, quoting both.
Then the actions for the final tier only. Tier 1: dates and baseline. Tier 2: the recurring claim, its providers, the source worth contacting and one ticket, only if the claim appears in at least two answers from two providers. Tier 3: the same plus what goes out today.
Ask me who owns escalation and on which channel; no tool knows who to wake up. Do not predict how much sentiment will recover.
Connect Qwairy to Claude, pull the exact signals in the workflow, and leave with an execution-ready output.
Audit whether AI engines accurately represent your brand. Compare AI-generated descriptions with your intended positioning, track narrative drift over time, and build a correction strategy.
Audit AI answers claim by claim against an approved source of truth, rank factual risks, trace cited evidence, plan corrections, and run a comparable re-test.
Audit the technical and content signals that support browser-capable AI agents, then build a structured, testable readiness roadmap with Qwairy MCP and Claude.