4. Measurement & Continuous Optimisation

Phase 4 runs from Month 2 as a monthly cycle. GEO is not a one-off exercise: AI models update continuously and the sources they answer from change, so visibility is created and preserved by regularly feeding and measuring the system built in Phases 1 to 3. This section specifies the query set under test, the monthly cycle, the metrics tracked, how gaps convert into work, and the monthly report. It measures observed assistant behaviour; it does not control it.

4.1 Query set under test

Measurement runs against a fixed baseline rather than an ad hoc set of prompts, so month-on-month movement is readable.

Starting point: The preliminary analysis Polenist carried out before this proposal found Aerion absent from all four assistants across the tested queries, with no Aerion record in the directories, media and reference sites that feed the models, and no AI-specific configuration on the website. That zero reading is the baseline the first monthly report is written against, so the first cycle measures movement rather than establishing the starting point from scratch.

4.1.1The target query matrix defined in the content engine (§3.1) becomes the measurement baseline. The clusters carry the same reference codes in both sections, so a measured result maps directly to the content that is meant to answer it.
4.1.2Every cycle is run on four platforms — ChatGPT, Claude, Gemini and Perplexity — because their answer sources and update behaviour differ, and a result on one is not evidence of a result on another.
4.1.3Queries are run in both content languages so the Turkish and English content bases are measured separately rather than assumed to behave alike.
4.1.4The baseline is held stable. Queries added later are recorded with the month they were introduced and reported alongside, rather than folded silently into the earlier figures.
4.1.5Runs are executed under comparable conditions each month: same wording, same platform surface, and no personalised session history that would make one month incomparable with the next.

4.2 Monthly cycle

One cycle per month, in four steps. The cycle starts in Month 2, once the first technical configuration and the first content set are live, and continues for as long as the monthly process runs.

Ref	Step	What happens
M-01	Test	The fixed query set is run across ChatGPT, Claude, Gemini and Perplexity under the same wording each month. Answers are captured verbatim, together with the date, the platform and any sources the answer cites.
M-02	Track	Each captured answer is scored against the tracked metrics: whether Aerion is mentioned, whether the stated information is accurate, and whether a source is shown. Results are compared with the previous month's run.
M-03	Identify gaps	Queries where Aerion is absent, described inaccurately or unattributed are traced back to their cause — a missing third-party source record, a missing content asset, or an asset the assistants are not reading.
M-04	Optimise	Each gap becomes a named item in the action plan, assigned to source work, content work or technical configuration and prioritised against the query clusters that matter most to Aerion. The items that sit inside the monthly scope are then carried out within the cycle rather than only recommended; the rest are quoted or escalated.

4.3 What is tracked

Six metrics, recorded per platform and per query cluster. They are observations of assistant output on the dates the cycle was run, not platform-reported statistics.

Ref	Metric	Definition
KPI-01	Mention rate	The share of tested queries in which Aerion appears in the assistant's answer. Reported per platform and in aggregate, and broken down by query cluster.
KPI-02	Information accuracy	Whether what the assistant states about Aerion matches the approved product narrative. Errors, omissions and outdated figures are recorded individually rather than scored as a single number.
KPI-03	Source attribution	Whether the answer shows a source at all, and whether that source is an Aerion-owned property or a third-party record built under Phase 2.
KPI-04	Cited source inventory	The specific sources the assistants draw on for the query set, listed per platform. This is what tells the source work which directories, media and reference sites are actually being read.
KPI-05	Narrative consistency	Whether the positioning language returned by the four platforms agrees with itself and with the published content base. Divergence between platforms is reported as a finding.
KPI-06	Cluster coverage	The proportion of clusters in the query matrix (§3.1) with at least one answer in which Aerion is mentioned, tracked month on month.

Unit of record: One query, on one platform, in one language, on one date. Every metric above is derived from that record, and the captured answers are kept so a figure can be traced back to the text it came from.
Comparison: Each month is reported against the previous month and against the first cycle, so both movement and cumulative position are visible.

4.4 Gap identification and optimisation actions

A measurement result is only useful if it changes what is produced next. Each finding is classified by cause, converted into named work, and — where it falls inside the monthly scope — carried out within the cycle rather than left as a recommendation.

Source gap — no third-party record exists that an assistant could read on the subject of the query. Routed to the source and authority work.
Content gap — no published asset answers the query in a citable form. Routed to the content engine and placed in the content calendar.
Technical gap — the asset exists but is not reachable, structured or marked up in a way AI systems can use. Routed back to the technical configuration.
Accuracy gap — the assistants answer, but with information that is wrong or out of date. Handled by correcting the underlying source or asset rather than by adding new material.

4.4.1Findings are prioritised by the commercial weight of the cluster they sit in, not by how easy they are to close.
4.4.2Content gaps enter the content calendar as specified items — subject, format and the query cluster they are meant to answer — so the content engine is fed by measurement rather than by assumption.
4.4.3Source gaps become registration, correction or placement tasks against the source asset map, including corrections to records that already exist but carry outdated information.
4.4.4Technical gaps return to the technical configuration as discrete fixes, which are re-tested in the following cycle.
4.4.5Items that depend on an Aerion decision, approval or external budget are flagged as blocked in the plan rather than carried silently.

Boundary — What the monthly scope carries out

The monthly process executes maintenance-scale optimisation: corrections to published assets and existing source records, schema and configuration fixes, calendar re-sequencing, and the content volume agreed for the month. Work that amounts to a new work package — a further language, a new content programme, or a registration set beyond the source asset map at §2.5 — is identified in the plan and quoted separately rather than absorbed into the monthly fee.

4.5 Reporting

Each cycle closes with one document: the monthly AI visibility report, issued with the action plan for the month ahead.

Cadence: One report per month, from Month 2 onward, for as long as the monthly process runs.
Contents: The metric table for the month, per platform and per cluster; movement against the previous month; the list of sources the assistants actually cited; the classified gap findings; and the action plan for the next cycle.
Evidence: The captured answers behind the figures are retained and referenced, so any reported number can be checked against the text that produced it.
Recipient: Issued to the Aerion point of contact, who circulates it internally. Distribution beyond that is an Aerion decision.
Deliverable: Monthly AI visibility report + action plan.

Boundary — What measurement can and cannot show

AI assistant outputs are non-deterministic and controlled by the platform vendors. The same query can return different answers on different runs, and a vendor can change its model, its retrieval behaviour or its source selection without notice. Measurement reports observed behaviour across a defined query set at a point in time. No ranking, placement or answer content can be guaranteed, and no figure in a report should be read as a commitment about the next one.